The use of clinical data from electronic medical records (EMRs) for clinical research and for evaluation of quality of care requires an extraction process. Many efforts have failed because the extracted data seemed to be unstructured, incomplete and ridden by errors. We have developed and tested a concept of extracting semi-structured EMRs (Journal III, Profdoc) data from 776 diabetes patients in a general practice clinic over a 5 year period. We used standard database management techniques commonly applied in clinical research in the pharmaceutical industry to clean up the data and make the data available for statistical analysis. The key problem was difficulties locating the data, as no standard way to enter the data in the EMR system was reinforced. Furthermore, no built-in edit checks to facilitate data entry were available. Laboratory, drug information and diagnostic data could be used directly while other data such as vital signs required much work to locate and become useful.