Researchers from the Massachusetts Institute of Technology have developed a new system that uses algorithms to more accurately interpret physicians' freeform notes in electronic health records, FierceHealthIT reports.
Next week, the researchers will present the system at the American Medical Informatics Association's annual symposium in Chicago.
Researchers from MIT's Computer Science and Artificial Intelligence Laboratory adapted algorithms from a type of research known as topic modeling (Hall, FierceHealthIT, 10/31). Topic modeling seeks to automatically identify document topics by drawing correlations between words that are prominently featured.
The researchers then developed an algorithm that identifies relationships between words while also drawing correlations between words and other textual features, such as syntax.
According to the researchers, the system is 75% accurate in distinguishing the meaning of words that could be interpreted in multiple ways.
Researchers say the algorithm offers a more accurate approach to "word-sense disambiguation" than previous methods.
They also say that the system could be used to analyze large bodies of text without human supervision.
Researchers plan to augment the system with listings from the Unified Medical Language System, a large thesaurus of medical terms compiled by NIH.
They also might incorporate:
- Additional syntactic and semantic features; and
- Word associations developed by NIH's Medical Subject Headings paper-classification scheme.
Hongfang Liu -- an associate professor of medical informatics at the Mayo Clinic -- said, "About 80% of clinical information is buried in clinical notes," adding, "A lot of words or phrases are ambiguous there."
Liu added that the new system potentially could "be used in production-scale natural language processing systems" (Hardesty, MIT News, 10/31).