Toward Automated Detection of Biased Social Signals from the Content of Clinical Conversations