TL;DR
This paper introduces TAXN, a novel pipeline for extracting medical entities from clinical texts in multiple languages. By combining translation, word alignment, and entity extraction, TAXN effectively links entities to English-based medical terminologies like UMLS, even when the source text is not in English. This approach significantly enriches multilingual clinical data for secondary use.
What Problem Does This Paper Address?
A vast amount of valuable information in electronic health records (EHRs) remains locked in unstructured text, making it difficult to analyze and utilize for research or clinical decision-making. While numerous tools exist for extracting information from English clinical texts, a significant challenge arises when dealing with texts in other languages. Medical terminologies are also far richer and more comprehensive in English (e.g., UMLS) compared to many other languages. Directly translating these texts and then extracting entities can break the crucial link between the original text and the extracted information. This paper tackles the problem of enabling robust, multilingual clinical text analysis by proposing a method that preserves the connection between the original text and the extracted entities, even when leveraging English-centric resources.
Methods
The core of the proposed solution is the TAXN (Translate Align Extract Normalize) pipeline. This pipeline operates in several key stages:
- Translation: Clinical texts in a non-English language (e.g., French) are translated into English.
- Alignment: A crucial step that establishes a link between the translated English text and the original text. This is achieved through word alignment techniques, ensuring that the origin of each translated word can be traced back.
- Entity Extraction: Named Entity Recognition (NER) is performed on the translated text to identify medical entities.
- Normalization: The extracted entities are then normalized to a standard medical terminology, specifically targeting the Unified Medical Language System (UMLS) in English.
The authors implemented this pipeline using an open-source Python library called Medkit. Medkit is designed for building and deploying data processing pipelines in healthcare, with a focus on data provenance and non-destructive processing. The TAXN pipeline within Medkit is constructed by chaining together several processing modules: a rule-based sentence tokenizer, a UMLS matcher (using QuickUMLS), a deep-learning translation and alignment module (Opus MT and awesome Align), a pre-trained BERT-based NER module (from HuggingFace Hub), and a UMLS normalizer (CODER).
Datasets
The study utilized a corpus of 70 French articles related to long COVID, manually collected from the web. This dataset was used for both the extraction of concepts and for evaluating the quality of the proposed synonyms. The paper does not mention any specific training datasets for the models used, implying reliance on publicly available pre-trained models.
Key Insights and Ideas
- Bridging the Multilingual Gap: The TAXN approach effectively overcomes the limitations of language barriers in clinical text analysis by leveraging the richness of English medical terminologies.
- Preserving Text-Entity Linkage: The translation-alignment step is critical, ensuring that extracted entities can be directly mapped back to their original context in the source text, which is vital for interpretability and validation.
- Enriching Phenotypic Dictionaries: The method proved effective in enriching a phenotypic dictionary for post-acute sequelae in COVID-19 (PASC), demonstrating its practical utility.
- Leveraging Pre-trained Models: The study highlights the feasibility of using freely available, pre-trained deep-learning models for complex NLP tasks in the medical domain without requiring specific training on domain-specific data.
Key Results
- The TAXN pipeline was able to extract a significant number of unique concepts, with the "ner-entities" and "entities-en" methods yielding 437 and 417 unique concepts, respectively, from the French corpus.
- The proposed synonyms, when manually evaluated, showed a high overall median accuracy of 0.81. The dictionary-based approach ("entities-en") achieved a slightly higher median accuracy of 0.85 compared to the NER approach (0.75).
- The system successfully enriched 356 terms with at least one new synonym and 113 terms with more than one synonym.
- A qualitative analysis showed that the NER approach could extract more diverse synonyms.
Limitations and Future Work
- Human-in-the-Loop Requirement: While the accuracy of proposed synonyms is high, a human is still required to validate them, limiting the potential for full automation at scale.
- Computational Cost: The pipeline involves multiple transformer-based models, making it computationally expensive and potentially challenging for production deployment.
- Future Work: The authors suggest exploring more "frugal" models or combining different pipeline steps into a single model to reduce computational overhead. Investigating methods to further automate the validation process is also a promising direction.
Summary updated