Extracting Structured Data from Physician-Patient Conversations By Predicting Noteworthy Utterances