A method for automatic extraction of clinical temporal information would be of significant practical importance for deep medical language understanding, and a key to creating many successful applications, such as medical decision making, medical question and answering, etc. This paper proposes a rich statistical model for extracting temporal information from an extremely noisy clinical corpus. Besides the common linguistic, contextual and semantic features, the highly restricted training sample expansion and the structure distance between the temporal expression & related event expressions are also integrated into a supervised machine-learning approach. The learning method produces almost 80% F- score in the extraction of five temporal classes, and nearly 75% F-score in identifying temporally related events. This process has been integrated into the document-processing component of an implemented clinical question answering system that focuses on answering patient-specific questions (See demonstration at http://hitrl.cs.usyd.edu.au/ICNS/).