MLM-based typographical error correction of unstructured medical texts for named entity recognition.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Electronic Health Records (EHRs) contain valuable research data but are hindered by frequent typographical errors.
• Extracting structured information from unstructured medical text is challenging due to data quality issues.
• Existing methods for correcting typos in medical records are limited, necessitating improved approaches.

Purpose of the Study:

• To develop and evaluate a novel typo correction model for unstructured medical texts.
• To enhance the accuracy of information extraction from surgical pathology records and other medical data.
• To overcome the limitations of current methods in handling real-world medical data errors.

Main Methods:

• A context-aware typo correction model based on the Masked Language Model (MLM) was developed.
• A word dictionary derived from PubMed abstracts was integrated into the model.
• Fine-tuning of a pre-trained BERT model followed by deep learning-based Named Entity Recognition (NER) was performed on corrected data.

Main Results:

• The proposed typo correction model demonstrated superior performance compared to the existing SymSpell model.
• The model achieved an approximate 5% and 9% F1-score improvement on the NCBI-disease and surgical pathology datasets, respectively.
• Named Entity Recognition (NER) performance increased by 2% on the NCBI-disease dataset and showed a significant 25% improvement on surgical pathology records post-correction.

Conclusions:

• Typographical errors in unstructured medical text significantly degrade natural language processing task performance.
• The proposed context-aware typo correction model offers a robust solution for real-world medical data analysis.
• This methodology effectively addresses data quality challenges, improving the reliability of information extraction from medical records.