In recent years, the widespread adoption of Hospital Information Systems is enabling hospitals to measure and record an ever-growing volume and variety of patient and process-related data. In such context, analytics are emerging as suitable tools and methods for extracting and analyzing such data and for providing useful insights to assist decision-making. The Emergency Department (ED) is one of the functional areas that could profit from the implementation of such tools and methods. However, the complex and dynamic nature of EDs makes the application of analytics a very challenging task which is attracting increasing attention from both academic researchers and practitioners. This work represents a first attempt to demonstrate the suitability of predictive analytics in ED environment. Specifically, we aim at developing a multi-level classification model to predict patients’ length of stay (LOS), by exploiting historical information about ED patients and the advanced machine learning techniques. Given the high variability affecting patient LOS, we determined a set of candidate predictors from the data, including patients’ characteristics, temporal factors and, system-based factors (e.g., the system workload, the abandon rate and the probability of treatments’ execution like radiology, laboratory and consultancy). Preliminary results show that the multi-level model achieves promising values of accuracy, outperforming the single-level classification model. It allows the early identification of patients likely to experience a long LOS in ED. Such patients may require a dedicated monitoring by ED service providers who should take appropriate actions to shorten their stays. Accordingly, predicting LOS can help ED management to dynamically monitor the crowding level of the system and make informed decisions about resource allocation.