Nowadays, the concept of education for all is gaining momentum thanks to the widespread use of e-learning systems around the world. The use of e-learning systems consists in providing learning content via the Internet to physically dispersed learners. The main challenge in this regard is the high fail rate particularly among k-12 learners who are our case study. Therefore, we established an in-depth methodology based on machine learning models whose objectives are the early prediction of at-risk learners and the diagnosis of learning problems. Going through this methodology was of a great importance thus it started by identifying the most relevant learning indicators among performance, engagement, regularity and reactivity. This, then, led us to extract and select the adequate learning features that reflect the activity of an online learner. For the modeling part of this methodology, we apply machine learning models among k-nearest neighbors (K-nn), Support Vector Machine (SVM), Random Forest and Decision tree on a real data sample of 1361 k-12 learners. The evaluation step consists in comparing the ability of each model to correctly identify the class of learners at-risk of failure using both accuracy and False Positive Rate (FPR) measures.