The unprecedented long-term online learning caused by COVID-19 has increased stress symptoms among students. The e-learning system reduces communications between teachers and students, making it difficult to observe student’s mental issues and learning performance. This study aims to develop a non-intrusive method that can simultaneously monitor stress states and cognitive performance of student in the scenario of online education. Forty-three participants were recruited to perform a computer-based reading task under stressful and non-stressful conditions, and their eye-movement data were recorded. A tree ensemble machine learning model, named LightGBM (Light Gradient Boosting Machine), was utilized to predict stress states and reading performance of students with an accuracy of 0.825 and 0.793, respectively. An interpretable model, SHAP (SHapley Additive exPlanation), was used to identify the most important eye-movement indicators and their effects on stress and reading performance. The proposed model can serve as a foundation for further development of user-centred services in e-learning system.