Extracting Diagnoses from Discharge Summaries

We have developed a program for extracting the diagnoses and procedures from the past medical history and discharge diagnoses in the discharge summary of a case and coding these using SNOMED-CT in the UMLS. The program uses a limited amount of natural language processing. Rather, it makes use of the relatively standard structure of the discharge summary, a small dictionary to divide the text into phrases, and the extensive collection of phrases for concepts in the UMLS to do the coding. With this approach the program finds 240 of 250 desired concepts with 19 false positives in 23 discharge summaries.