The potential impact of prediction methods is particularly relevant to emergency department (ED) settings which are characterized by complex and challenging conditions. Repeat ED visits are one such example of potential over-utilization over a longitudinal period that can in some cases be predicted. We tracked the ED revisit risk over time using a hidden Markov model (HMM) based on adult ED ten years encounters at a single US urban hospital. Given the HMM states, we performed prediction for future ED revisits. We applied Decision Jungle Trees and Logistic Regression prediction models. The data were combined from distributed sources (e.g. EHRs and HIE). The results show that integrating a pre-analysis of the diverse patients using latent class models and applying them to predictive classifiers performed well. The performance was better than without using pre-analysis of HMM. These findings suggest that one potential approach to improved risk prediction is to leverage the longitudinal nature of health care delivery.