Inferred joint multigram models for medical term normalization according to ICD.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Electronic Health Records (EHRs) utilize spontaneous natural language, often deviating from standardized terminologies.
• Discrepancies between natural language terms in EHRs and standard classifications like the International Classification of Diseases (ICD) hinder information exchange.

Purpose of the Study:

• To develop a system for normalizing diagnostic terms from EHRs into the standard ICD framework.
• To improve information retrieval and exchange by bridging the gap between spontaneous and standard medical terminology.

Main Methods:

• Weighted Finite-State Transducers (WFSTs) were employed for diagnostic term normalization.
• A similarity metric was implemented to enhance the matching of spontaneous and standard medical terms.
• WFSTs were trained on sample data to translate natural language representations into standardized ones.

Main Results:

• Only 7.71% of 2850 spontaneous diagnostic terms directly matched standard ICD forms.
• The WFST-based system achieved a Mean Reciprocal Rank of 0.68 for matching spontaneous to ICD terms.
• This indicates the correct ICD code was typically found within the top two normalized candidates.

Conclusions:

• High-performance medical term normalization was achieved using WFSTs.
• Direct matching of spontaneous terms to standard lexicons yielded unsatisfactory results.
• WFST-based normalized hypothesis generation effectively addressed the disparity between spontaneous and standard medical language.