Cancer stage, an important element in prognosis, is often documented in narrative form in medical records, which leads to time-consuming abstraction by tumor registry personnel and other secondary users of these records.
Jeremy Warner, M.D., M.S., and colleagues developed a computerized natural language processing algorithm to extract cancer stages from electronic medical records. Records from 2,323 lung cancer patients were used to train and test the algorithm, and accuracy was evaluated as records from successively longer periods were processed — one week after diagnosis, two weeks, and so on up to 26 weeks.
The researchers report in the Journal of Oncology Practice that as new information is obtained and as multiple clinicians work to forge consensus, medical records, especially those from early in the course of treatment, may reflect ambiguity regarding stage.
Accuracy of the algorithm surpassed 50 percent within six weeks of diagnosis, and within 14 weeks reached 90 percent, despite that more than 80 percent of the records were found to harbor conflicting information about stage.
Warner was joined in the study by Mia Levy, M.D., Ph.D., and Michael Neuss, M.D.