Statistical methods that can support text retrieval are becoming an increasing focus of medical informatics activities. We overview our adaptation of existing knowledge sources to create pseudo-documents for concept based latent semantic indexing. Experience demonstrated this tack of limited practical value, since retrieval performance was invariably unsatisfactory. We discovered this was due in part to the introduction of a vocabulary gap between the queries and the cases we sought to retrieve. In part to address this problem, and to avail our large body of humanly coded text as a knowledge source, we developed a least squares fit alternative for the computer assisted indexing and retrieval of biomedical texts. This technique demonstrates equivalent or superior retrieval performance when compared to all other textual retrieval techniques. It does not depend upon elaborate knowledge bases, lexicons, or thesauri. It is a promising technique for classifying and retrieving the large volumes of clinical text.