Combining voice recognition and automatic indexing of medical reports

Medical records have been evolving from the traditional paper-based records to digital ones, from the method of dictating reports and transcription to voice recognition systems. The transition to digital operations will not be complete until we have the ability to combine voice recognition with automated indexing of texts. This paper introduces the methods we used to evaluate existing voice recognition software programs and presents NOMINDEX, a system that turns a medical text into MeSH codes, using the French ADM lexical database. Those systems were applied to 28 patient discharge summaries in French, produced after a coronarography, and extracted from the MENELAS corpus of texts. Using the best configuration for voice recognition, the rate of accurate recognition exceeds 98 percent. Among the indexing concepts assigned by NOMINDEX, 25 percent were not pertinent and 12 percent of the relevant concepts were missing. Most errors were related to confusion between common language and medical language, and to the coverage of the ADM lexical database. Best results would be expected with a more comprehensive lexical resource In addition, only 3 percent of the errors generated by inadequate voice recognition that remained in the configuration that performed better, impacted on automatic indexing by NOMINDEX.