This paper presents a comprehensive study of learning assessment, delving into the concept of item difficulty and learner perception. It addresses two critical dimensions: the methodologies employed, particularly data-driven approaches, and the necessary data for this analysis. Traditional difficulty estimation methods focus on question content or student performance. Recent studies suggest using machine learning and natural language processing to predict question difficulty. These models are subject-specific and often overlook individual student differences, limiting their wider application. The work aims to examine data of real-world testing scenarii, so that assembling and building a rich and diverse dataset. It offers valuable insights into the factors influencing item difficulty by giving the maximum amount of information considering the test and the student. It presents experiments to build and train predictive machine learning models for difficulty prediction. At the end, thanks to experiments, we can show a nuanced understanding of the assessment challenge and lay the groundwork for incorporating psychological factors into difficulty estimation as a subsequent phase.