Assessing quality and agreement of structured data in automatic versus manual abstraction of the electronic health record for a clinical epidemiology study