Determining cause of death in 45,564 autopsy reports

It has been demonstrated that death certificates do not accurately record the actual cause of death in up to one-fourth of cases, as determined from subsequent autopsy findings. The purpose of this study was to explore the use of natural language autopsy data bases as an automated quality assurance mechanism. We translated the account of the major process leading to death, or the primary diagnosis, from all 45,564 narrative autopsy reports obtained at The Johns Hopkins Hospital between May 28, 1889, and June 30, 1987, into the hierarchical system of Medical Subject Heading (MeSH) titles. We obtained a total of 125,772 MeSH title translations, 1,563 of them distinct (average 2.8 per case), ranging in frequency from 6,029 occurrences of LUNG to 1 occurrence apiece of 357 MeSH titles. The natural-language-to-MeSH translations showed expected trends over the past century: fewer infectious diseases; more cardiovascular and neoplastic disease among adults; and more respiratory diseases and congenital malformations in the pediatric age group. The greater availability of autopsy documents in electronic form should increase the value of this resource for quality assurance.