Lessons Extracting Diseases from Discharge Summaries

We developed a program to extract diseases and procedures from discharge summaries and have applied this program to 96 cases annotated by physicians. We compared the concepts extracted by the program to those extracted by the annotators. The program extracts 93% of the desired concepts including some more specific than the annotators. Concepts were missed because phrases were ambiguous, phrases were missing words or were separated, or deduction was needed, among other reasons. The false positives were either insignificant findings, ambiguous phrases, or did not apply to the patient now. The analysis shows that extraction of medical concepts from discharge summaries with limited natural language processing and no domain inference is effective with still more potential.