Natural Language Processing and Web Tools for Mapping Units from ClinicalTrials.Gov

摘要

Summary data from some clinical trials is now becoming available in electronic format due to US public law through the NIH NLM database ClinicalTrials.Gov. The database currently holds over a quarter of a million trials with about 30K trials with result records. However, despite the great work of the team that developed this fast growing database, the data held in it is far from standardized and using it requires effort, especially for machines. The difficulty arises from the fact that data entry to this database is manual from multiple external sources, mostly textual, and somewhat permissive. Although entered data goes through a review process, the review is human and therefore sometimes forgiving. For machine comprehension, definition of units is essentials so numbers held in the datab ase will make sense. There are currently over 20K units for 30K clinical trials with results and many of the units are synonyms and some are even errors. Even CDISC units that are in a good level of standardization need normalization and enhancement. This presentation will discuss recent advances in the effort to standardize the medical units. Specifically Natural Language Processing (NLP) techniques are used alongside other Machine Learning methods to cluster similar units together. This allows a human to inspect and standardize the units more efficiently. In this presentation the NLP techniques used will be discussed in details as well as the clustering techniques. Similar techniques merged with web tools will be helpful for future analysis of other textual fields within the fast growing database. To move the standardization effort forward a web tool was created to allow humans to classify the units. Multiple users can see the units in each cluster and the machine suggestions and classify those. The web portal is accessible through: ClinicalUnitMapping.com.