In NeuroIS, cognitive state inference provides rich insight into users’ cognitive experiences during information technology use. However, inferences drawn from electroencephalography (EEG) - one of the most common methods used in NeuroIS - face critical challenges in real-world applications due to label scarcity and cross-task variability. This work introduces a self-supervised learning framework that combines masked prediction and contrastive learning to learn robust EEG representations from unlabeled data. Pre-trained on both controlled (N-back) and naturalistic (MATB-II) tasks, our approach employs hybrid fine-tuning, leveraging full labels from N-back and sparse MATB-II annotations to enhance cross-task generalization. Evaluations demonstrate a 68.2% classification accuracy on MATB-II. This work establishes a framework for label-scarce EEG analysis in NeuroIS, advancing mental state measurement in ecologically valid settings where annotations are limited.