wirm-logo2.gif (1722 bytes)

Fathom: Natural Language Processing of Clinical Reports


Fathom is part of a natural language processing (NLP) system being developed by the UCLA Telemedicine group.  The system processes free-text clinical radiology reports and automatically generates structured records about the findings contained therein.  Structured reports are clinically useful for decision support and outcomes research.  The NLP system uses various statistical and machine learning methods to process the records and identify the properties, locations and diagnostic interpretations of each finding mentioned in the text.   Fathom is a Wirm-based application that supports the UCLA researchers as they test the effectiveness of various NLP algorithms.  In order to evaluate the recall and precision of the NLP engine, the results must be compared to thousands of hand-coded candidates.  Fathom reads the candidates from the NLP engine and presents them in web forms to the domain experts, who are able to submit their judgments from any location.  In addition, Fathom provides a Patient Record Manager that imports records from the hospital information system, allowing authorized users to retrieve records by patient demographics or by experiment corpus.   Fathom allows searching over the raw records or the intermediate data structures used by the various stages of processing.  For example, the syntactic parser splits the records into sentences, which are each assigned a sequential identifier for reference by the semantic interpreter.   The various processing stages express their output as XML records, which are read by the Patient Record Manager and made available for structured browsing.  Perl’s built-in text parsing abilities greatly facilitated the process of importing these records. 


Because the system contains sensitive clinical content, we have restricted access to the interface.  To request access to the system, please email: info@workhost.com.