Strategies for searching medical natural language text. Distribution of words in the anatomic diagnoses of 7000 autopsy subjects.

摘要

Computerized indexing and retrieval of medical records is increasingly important; but the use of natural language versus coded languages (SNOP, SNOMED) for this purpose remains controversial. In an effort to develop search strategies for natural language text, the authors examined the anatomic diagnosis reports by computer for 7000 consecutive autopsy subjects spanning a 13-year period at The Johns Hopkins Hospital. There were 923,657 words, 11,642 of them distinct. The authors observed an average of 1052 keystrokes, 28 lines, and 131 words per autopsy report, with an average 4.6 words per line and 7.0 letters per word. The entire text file represented 921 hours of secretarial effort. Words ranged in frequency from 33,959 occurrences of "and" to one occurrence for each of 3398 different words. Searches for rare diseases with unique names or for representative examples of common diseases were most readily performed with the use of computer-printed key word in context (KWIC) books. For uncommon diseases designated by commonly used terms (such as "cystic fibrosis"), needs were best served by a computerized search for logical combinations of key words. In an unbalanced word distribution, each conjunction (logical and) search should be performed in ascending order of word frequency; but each alternation (logical inclusive or) search should be performed in descending order of word frequency. Natural language text searches will assume a larger role in medical records analysis as the labor-intensive procedure of translation into a coded language becomes more costly, compared with the computer-intensive procedure of text searching.