Condition-associated EEG signatures in N-back tasks: temporal dynamics, classification, and generalization

Mental workload (MWL) monitoring with electroencephalography (EEG) is promising for adaptive systems, but practical use requires models that remain reliable across time, tasks, and users with feasible sensor layouts. This study investigated condition-associated EEG signatures in 15 healthy adults performing two complementary N-back paradigms: an N-back level manipulation (N-LEVELS: 1-back, 2-back, 3-back) and a stimulus-pacing manipulation at fixed 2-back (N-SPEED: slow, medium, fast). EEG was recorded using a compact 8-channel setup. Spectral bandpower features, composite workload indices, and N2/P3b ERP markers were analyzed. Linear classifiers were evaluated using within-subject TRAIN → TEST and leave-one-subject-out (LOSO) protocols, together with analyses of observation duration, time-on-task dynamics, temporal generalization, cross-paradigm transfer, and exploratory ordinal regression. N-LEVELS and N-SPEED produced distinct behavioral and subjective profiles: increasing N-back level was associated with greater task demand, whereas faster N-SPEED conditions showed greater perceived temporal demand and reduced response-time availability. Exploratory locally corrected spectral analyses suggested more frequent beta-band and engagement-index modulation in N-LEVELS and more frequent posterior/parietal alpha suppression and theta/alpha-related modulation in N-SPEED. ERP analyses showed reduced P3b amplitude under higher N-back levels and faster pacing. Classification performance was modest and context-dependent: N-SPEED three-class decoding was more consistent than N-LEVELS, and all three N-SPEED pairwise contrasts survived correction under both WITHIN and LOSO evaluation, whereas in N-LEVELS only the 1-back vs. 3-back contrast survived under LOSO. Post hoc artifact-sensitivity analyses supported the principal P3b findings and largely preserved reference-decoding and alpha time-on-task patterns. After delta-band removal, neither three-class LOSO result nor the N-SPEED slow-vs.-medium and medium-vs.-fast LOSO contrasts survived correction; N-SPEED slow vs. medium under LOSO also did not survive either high-pass variant. Exploratory ordinal regression showed a clearer ordered relationship for N-SPEED than N-LEVELS. Time-on-task analyses revealed selective non-stationarity, mainly progressive alpha build-up, whereas cross-paradigm transfer did not survive corrected inference. These findings indicate that compact EEG recordings contain condition-associated information, but the small sample, deterministic block order, modest decoding performance, and limited cross-paradigm generalization constrain causal attribution and applicability beyond the present protocol.