An experiment comparing lexical and statistical methods for extracting MeSH terms from clinical free text

Area of Science:

• Medical Informatics
• Natural Language Processing
• Information Retrieval

Background:

• Electronic medical records (EMRs) contain valuable clinical information.
• Efficiently searching EMRs for specific patient data is crucial for healthcare.
• Developing automated systems to index clinical text can improve information retrieval.

Purpose of the Study:

• To develop and comparatively evaluate PostDoc (lexical) and Pindex (statistical) indexing systems.
• To assess the performance of a hybrid system combining PostDoc and Pindex.
• To explore methods for assisting healthcare personnel in constructing patient-specific Medline searches from EMRs.

Main Methods:

• Tested PostDoc, Pindex, and a hybrid system on clinical records (radiology, pathology, discharge summaries).
• Identified relevant concepts within the clinical text for Medline search formulation.
• Evaluated system performance based on the proportion of relevant concepts captured in the suggested MeSH terms via blinded assessment.

Main Results:

• PostDoc identified ~40% of relevant concepts, outputting ~19 MeSH terms per report.
• Pindex identified ~45% of relevant concepts, outputting ~57 MeSH terms per report.
• The hybrid system captured ~66% of relevant concepts, outputting ~71 MeSH terms per report.

Conclusions:

• PostDoc and Pindex outputs are complementary for capturing MeSH terms from clinical text.
• The hybrid system demonstrated superior performance in concept capture.
• Future work may involve using UMLS semantic types to refine search results and improve clinical relevance.