Term-BLAST-like alignment tool for concept recognition in noisy clinical texts

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Concept recognition (CR) methods are often evaluated on standard medical literature.
• Electronic Health Records (EHRs) present unique challenges for CR due to nonstandard text like misspellings and abbreviations.

Purpose of the Study:

• To develop and evaluate a novel method for concept recognition in clinical texts, specifically addressing challenges in EHR data.
• To improve the accuracy and efficiency of entity linkage in biomedical and clinical text mining.

Main Methods:

• A Term-BLAST-like alignment tool (TBLAT) was developed, inspired by the BLAST algorithm for biosequence alignment.
• TBLAT uses k-mer counts and a gold standard corpus of typographical errors from clinical notes to score potential concept matches.
• The method was experimentally compared against five established tools using scientific publications and EHR records.

Main Results:

• TBLAT demonstrated a 10% increase in recall on scientific publications compared to existing methods.
• TBLAT achieved a significant 20% increase in recall on EHR records, highlighting its effectiveness on nonstandard clinical text.
• The tool can be utilized independently or as a supplementary component to enhance current CR approaches.

Conclusions:

• The Term-BLAST-like alignment tool (TBLAT) offers a substantial improvement for concept recognition in clinical text, particularly in challenging EHR environments.
• TBLAT's sequence alignment-inspired approach effectively handles variations and errors common in clinical notes, enhancing entity linkage.
• The Fenominal Java library provides an accessible implementation of TBLAT for Human Phenotype Ontology term recognition.