Surgery is a highly critical and costly procedure, and there is an imperative need to improve the efficiency in surgical wards. Analyzing surgical patient flow and predicting cycle times of different peri-operative phases can help improve the scheduling and management of surgeries. In this paper, we propose a novel approach to mining temporal patterns of surgical patient flow with the use of Bayesian belief networks. We present and compare three classes of probabilistic models and we evaluate them with respect to predicting cycle times of individual phases of patient flow. The results of this study support previous work that surgical times are log-normally distributed. We also show that the inclusion of a clustering pre-processing step improves the performance of our models considerably.