A UMLS-based spell checker for natural language processing in vaccine safety

Area of Science:

• Natural Language Processing (NLP)
• Public Health Surveillance
• Patient Safety Informatics

Background:

• Patient safety is a key healthcare goal in the U.S.
• Detecting vaccine adverse events is crucial for public health.
• Free-text reports from immunization surveillance systems require NLP for analysis.

Purpose of the Study:

• To develop a Unified Medical Language System (UMLS)-based spelling error correction tool.
• To serve as a foundational step in the NLP pipeline for adverse events following immunization (AEFI) reports.
• To improve the extraction of UMLS concepts from AEFI text data.

Main Methods:

• Developed spell-checking algorithms using open-source tools and de-identified AEFI reports.
• Implemented a four-step process: error detection, word list generation, disambiguation, and correction.
• Expanded abbreviated clinical terms and shortcuts before spell checking.

Main Results:

• Trained the spell checker on 12,056 words and tested on 8,131 words.
• Achieved 74% sensitivity, 100% specificity, and 47% positive predictive value (PPV).
• The prototype demonstrated superior specificity compared to existing tools.

Conclusions:

• A prototype UMLS-based spell checker for AEFI reports was successfully created.
• The tool utilized UMLS Specialist Lexicon and WordNet for dictionary terms.
• Methods can be adapted for text cleaning in other specialized domains.