Detecting the student internal state during learning is a key construct in educational environment and particularly in Intelligent Tutoring Systems (ITS). Students' uncertainty is of primary interest as it is deeply rooted in the process of knowledge construction. In this paper we propose a new sensor-based multimodal approach to model users' uncertainty from their affective reactions and cognitive and personal characteristics. An experimental protocol was conducted to record participants' brain activity and physiological signals while they interacted with a computer-based problem solving system and self-reported their perceived level of uncertainty during the tasks. We study key indicators from affective reactions, trait-questionnaire responses, and individual differences that are related to uncertainty states. Then we develop models to automatically predict levels of uncertainty using machine learning techniques. Evidence indicated that students' uncertainty is associated to their mental and emotional reactions. Personal characteristics such as gender, skill level, and personality traits also showed a priori tendencies to be more or less in particular uncertainty states. The SVM algorithm demonstrated the best accuracy results for classifying students' uncertainty levels. Our findings have implications for ITS seeking to continuously monitor users' internal states so they can ultimately provide efficient interventions to enhance learning.