TAXN: Translate Align Extract Normalize, a Multilingual Extraction Tool for Clinical Texts

Area of Science:

• Medical Informatics
• Natural Language Processing (NLP)
• Computational Linguistics

Background:

• Unstructured data constitutes approximately 80% of medical information in electronic health records.
• Limited availability of medical terminologies in non-English languages hinders the development and application of NLP tools in diverse linguistic contexts.

Purpose of the Study:

• To propose and implement a novel approach (TAXN) for leveraging English-based medical resources in other languages.
• To address the limitations of non-English NLP tools by utilizing translation, word alignment, entity extraction, and term normalization.
• To demonstrate the utility of the TAXN pipeline through a practical use-case in enriching a phenotypic dictionary for post-acute sequelae in COVID-19 (PASC).

Main Methods:

• Developed an extraction pipeline named TAXN, integrating translation, word alignment, entity extraction, and term normalization.
• Implemented the TAXN pipeline within the open-source library 'medkit'.
• Utilized freely available deep-learning models for the TAXN process.

Main Results:

• The TAXN approach proved efficient in identifying new synonyms for Unified Medical Language System (UMLS) terms.
• Enriched a phenotypic dictionary for PASC using a corpus of 70 French articles.
• Successfully validated at least one new synonym for 356 French terms.

Conclusions:

• The TAXN method effectively expands the utility of English medical resources for non-English languages, thereby improving NLP capabilities.
• This approach offers a viable solution for enhancing medical terminologies and dictionaries in multilingual settings.
• The open-source 'medkit' library and TAXN pipeline provide a valuable tool for researchers and developers in medical informatics and NLP.