ACXNet++ an attention augmented hybrid deep learning framework for cross task mental workload estimation from EEG neural manifolds

Abstract

Mental workload (MWL) estimation from electroencephalography (EEG) is central to human-computer interaction, aviation safety, and cognitive-state-aware systems, yet building models that generalize across tasks and across subjects without subject-specific calibration remains an open challenge. Abinaya and Dinakaran (2025) proposed ACXNet, a hybrid pipeline combining an autoencoder, a convolutional neural network (CNN), and an XGBoost classifier, evaluated on the STEW (Simultaneous Task EEG Workload) dataset, reporting 92.10% accuracy on the SIMKAP multitasking condition and 89.94% on a resting (No-Task) condition from a single held-out data split. This work presents ACXNet++, which inserts a multi-head self-attention layer between the CNN and the XGBoost classifier so that the model can learn, per input window, which spatial-temporal segments of the compressed EEG representation are most informative for the workload decision, rather than pooling them uniformly. Critically, this is not a numbers-only replication: the complete pipeline was implemented and executed end-to-end in this study against the real, publicly released STEW archive (48 subjects, 45 with usable subjective workload ratings), with every preprocessing, labeling, modeling, and evaluation choice documented explicitly, including places where the original paper under-specifies its protocol (the exact rating threshold used to binarize workload, and how its reported 64 handcrafted features are aggregated per channel). An initial evaluation using a single 70/15/15 subject split reproduced a serious and instructive failure mode: several feature configurations reached up to 100% training accuracy while producing test-set predictions that were confidently anti-correlated with ground truth (area under the ROC curve, AUC, as low as 0.033) a signature of subject-identity memorization rather than workload-pattern learning, made possible by the small number of held-out subjects. This was diagnosed and corrected by moving to subject-level stratified 5-fold group cross-validation and by regularizing the classifier more aggressively. Under this corrected protocol, adding attention improved mean cross-validated test accuracy on the SIMKAP condition from 80.73% (CNN alone) to 91.33% (CNN+Attention), and AUC from 0.910 to 0.937, a difference that was directionally consistent across all five folds (Wilcoxon signed-rank p = 0.0625, the smallest attainable p-value at this fold count). Gains on the resting condition were smaller and noisier. A six-configuration ablation study, aggregate cross-validated ROC and confusion-matrix analysis, attention-weight visualization, per-channel spectral analysis, and full per-fold result tables are reported. All findings, including negative and inconclusive ones, are reported transparently.