Recognizing and responding adequately to the learners' affective reactions has emerged as a key functionality for building a new generation of adaptive computing learning environments. This work presents a low intrusiveness hybrid model for inference of learning centered affective states. This model allows gathering information that indicates situations relevant to learning, such as, vicious cycle and engaged concentration. Promising results obtained in an experiment with students indicate the feasibility of this proposal and also support the presentation of alternatives for adapting the computational environment or implementing personalized pedagogical interventions.