Abstract
This review uniquely integrates open access dataset taxonomy with methodological trends in stress analysis, outlining the experimental framework and highlighting key gaps in reproducibility and FAIR compliance. In this context, stress induction methods, ground truth labeling approaches, open access datasets, computational advances, and current challenges and limitations are reported. A systematic review over the last decade (2014-2024) identified thirty-two open access affective datasets eligible for stress-related research, encompassing multimodal physiological signals, including electroencephalography (EEG), electrocardiography (ECG), electrodermal activity (EDA), and respiration (Resp), as well as behavioral measures, such as motion, audiovisual, and eye tracking data. Recent developments in signal analysis methods (2023-2025) highlight the predominance of multimodal fusion, advances in deep and self-supervised learning, personalized/adaptive models, and the growing adoption of explainable Artificial Intelligence, while machine learning approaches continue to hold a fundamental role. Despite these advances, several limitations and challenges remain, including heterogeneous experimental designs, sensor variability, limited demographic representation, data synchronization and labeling, and class imbalance. An effective experimental framework for stress research should integrate individual demographics and traits, reliable stressors, and high-quality physiological recordings within a well-defined and bias-controlled protocol, thereby producing reliable data to support and validate computational stress modeling. Continued progress in sensing, experimental standardization, and interpretable modeling is essential to produce reproducible, interpretable, and generalizable models of stress and emotions.