Identifying at-risk students as soon as possible is a challenge in educational institutions. Decreasing the time lag between identification and real at-risk state may significantly reduce the risk of failure or disengage. In small courses, their identification is relatively easy, but it is impractical on larger ones. Current Learning Management Systems store a large amount of data that could help to generate predictive models to early identification of students in online and blended learning. The contribution of this paper is twofold: First, a new adaptive predictive model is presented based only on students' grades specifically trained for each course. A deep analysis is performed in the whole institution to evaluate its performance accuracy. Second, an early warning system is developed, focusing on dashboards visualization for stakeholders (i.e., students and teachers) and an early feedback prediction system to intervene in the case of at-risk identification. The early warning system has been evaluated in a case study on a first-year undergraduate course in computer science. We show the accuracy of the correct identification of at-risk students, the students' appraisal, and the most common factors that lead to at-risk level.