A systematic review of explainable artificial intelligence for EEG based stress and mental state analysis
Abstract
Electroencephalography (EEG) has emerged as a key modality in neurotechnology for modeling stress and closely related mental states, owing to its non-invasive nature and high temporal resolution. Advances in machine learning have improved the performance of EEG-based predictive systems, yet the limited interpretability of many models continues to hinder their adoption in clinical and real-world settings. This systematic review synthesizes 46 studies to examine how Explainable Artificial Intelligence (XAI) is being applied to EEG-based models of stress and its most closely associated states, emotion, anxiety, depression, post-traumatic stress disorder, and cognitive workload or fatigue. Beyond narrative synthesis, every included study is appraised against a quality framework covering explainability method and scope, neurophysiological plausibility, validation protocol, leakage and class-imbalance risk, and ecological validity, and the XAI approaches are compared along local-versus-global, perturbation-versus-gradient, and post-hoc-versus-intrinsic dimensions. The analysis confirms a growing but still limited use of model-agnostic approaches such as SHAP and LIME. Three findings temper the field’s headline results. First, reported associations between EEG frequency bands and stress are frequently interpreted without adequate control for artifact contamination, recording reference, task and vigilance confounds, or the overlap between stress, arousal, workload, and fatigue. Second, studies reporting both subject-dependent and subject-independent validation show a large and systematic accuracy drop under the stricter protocol, and only a minority report subject-independent results at all. Third, no reviewed study formally tested whether its explanations are faithful or stable, and fewer than one in ten supplies evidence obtained outside a controlled laboratory setting. Together these findings indicate that a high reported accuracy is not, on its own, evidence of a trustworthy model, a faithful explanation, or a physiologically specific finding, and that this gap limits confidence in the transfer of reported accuracies to safety-critical contexts such as occupational health monitoring, clinical decision support, and human-robot interaction. The review concludes by setting out the reporting standards needed to close this gap: subject-independent validation, quantitative explanation-evaluation, and deployment-oriented rather than laboratory-inferred evidence.