Abstract
Mental workload is an interdisciplinary construct that significantly influences human performance, particularly in tasks requiring sustained attention and cognitive processing. Effective mental workload assessment is critical for preventing cognitive overload, which can lead to errors and reduced efficiency in high-stakes environments. The approach leverages topographic neural manifolds (spatial electrode arrangements) and temporal neural manifolds (time-series patterns) to capture comprehensive brain activity representations.Traditional methods rely on subjective reports or task performance, but physiological signals like EEG provide a more objective and continuous means of monitoring cognitive states. Therefore, this paper proposes a hybrid novel approach ACXNet which integrates autoencoder, CNN and XGBoost to learn features of EEG from an individual cross task performance without prior subject-specific calibration or task specific pre-labeled .training data. Utilizing the STEW (Simultaneous Task EEG Workload) dataset, containing recordings from 48 participants experiencing different levels of cognitive demands. Unsupervised feature extraction was carried out using an autoencoder. Subsequently, a CNN was employed to capture the spatial-temporal dependencies in the data, and XGBoost was utilized for efficient mental workload classification. This research adopts a binary classification approach to differentiate between low and high mental workload during SIMKAP and No task. The ACXNet model proposed in this study outperforms the existing methods with an average accuracy of 92.10% for SIMKAP task and 89.94% for No task condition. These findings show that ACXNet significantly improves the robustness and precision of mental workload estimation, providing a scalable solution adaptable to real-world applications, opening new avenues for the development of intelligent systems in human-computer interaction, healthcare, and beyond.