Unsupervised Clinical Language Translation

Wei-Hung Weng, Yu-An Chung, Peter Szolovits

TL;DR

This paper tackles the challenge of translating complex medical jargon in clinical notes into plain language that patients can understand, aiming to improve patient-clinician communication. They achieve this using a novel, fully unsupervised framework that combines representation learning, bilingual dictionary induction, and statistical machine translation, eliminating the need for expert-curated data.

What Problem Does This Paper Address?

Effective communication between doctors and patients is crucial for better health outcomes. However, clinical notes are often filled with professional jargon and abbreviations that are unintelligible to the average person. Statistics show that only a small percentage of adults are proficient in clinical language, leading to confusion, anxiety, and potentially poorer health decisions. While some approaches exist, they rely heavily on expert-curated dictionaries, which are difficult to scale and generalize. This paper addresses the critical need for an automated, unsupervised method to translate clinical language into consumer-understandable terms, making health information more accessible.

Methods

The authors propose a two-step, fully unsupervised framework:

  1. Word Translation: This step uses unsupervised word representation learning (specifically, the skip-gram model) to create embedding spaces for both professional and consumer language. They then employ Bilingual Dictionary Induction (BDI) techniques, such as MUSE and VecMap, to align these embedding spaces. A key innovation here is the use of "identical strings" (common words) found in both professional and consumer corpora as anchors to guide the alignment, minimizing the need for explicit bilingual dictionaries. They also explore the impact of subword information and dimensionality on embedding quality.
  2. Sentence Translation: This step leverages a statistical machine translation (SMT) framework. The BDI-aligned word embeddings are used to initialize the translation process. They then train language models (using KenLM) on monolingual corpora to refine the translations. Crucially, they employ iterative back-translation to generate parallel sentence pairs, allowing the SMT system to learn and improve over multiple iterations without requiring pre-existing parallel clinical data.

Datasets

  • MIMIC-III: De-identified clinical notes from ICU patients were used. Specifically, the "History of present illness" and "Brief hospital course" sections were used for the professional language corpus, while "Discharge instruction" and "Follow-up instruction" sections were used for the consumer language corpus.
  • MedlinePlus Corpus: This corpus of consumer-understandable health information was used to augment the MIMIC-III consumer language corpus, investigating its impact on BDI.
  • WMT English News Crawl Corpora: Used for training general language models for sentence translation.

Key Insights and Ideas

  • The Power of "Identical Strings" as Anchors: The paper highlights that using common words present in both professional and consumer texts as anchors for BDI significantly improves the alignment of embedding spaces, a crucial step for unsupervised translation. This bypasses the need for large, expert-curated bilingual dictionaries.
  • Trade-offs in Subword Information: While subword information can capture morphological patterns and improve word translation, it can sometimes lead to semantically incorrect translations. The paper shows that for sentence translation, embeddings without subword information can sometimes yield better correctness, as morphological errors can be corrected by language models.
  • Clinician-Led Evaluation is Crucial: The authors emphasize the importance of a two-step evaluation process: first, clinicians assess clinical correctness, and then both clinicians and non-clinicians evaluate readability. This approach mitigates biases and ensures both accuracy and understandability.
  • General vs. Specific Language Models: Using general language models (like WMT) can improve readability by simplifying jargon, but may sacrifice clinical correctness. Conversely, specific language models (like MIMIC-consumer) can maintain correctness but might be less accessible.

Key Results

  • Word Translation: Achieved a precision at 10 (P@10) of 0.827 on professional-to-consumer word translation. The MUSE BDI method generally outperformed VecMap. Using anchors for BDI was found to be critical.
  • Sentence Translation: Achieved mean opinion scores (MOS) of 4.10 for clinical correctness and 4.28 for layperson readability. Configuration F, which used 300-dimension word embeddings with anchors and the MIMIC-consumer corpus for language modeling, yielded the highest correctness scores. The supervised dictionary-based baseline had slightly better correctness but lower readability.

Limitations and Future Work

  • Correctness vs. Readability Trade-off: The unsupervised approach, while improving readability, sometimes leads to slightly lower clinical correctness compared to supervised methods. This oversimplification can be problematic in a clinical context.
  • Future Directions:
    • Leveraging existing dictionaries and clinical concept-level information to improve correctness.
  • Incorporating linguistic features for better control over language simplification.
  • Developing contextualized representations to enhance embedding quality.
  • Expanding the dataset of clinician-designed word pairs and deploying the framework as an online translator.
  • Investigating techniques to handle pronoun and negation ambiguity, and potentially using copy mechanisms or biomedical ontologies for specific terms.

Summary updated

Abstract

As patients' access to their doctors' clinical notes becomes common, translating professional, clinical jargon to layperson-understandable language is essential to improve patient-clinician communication. Such translation yields better clinical outcomes by enhancing patients' understanding of their own health conditions, and thus improving patients' involvement in their own care. Existing research has used dictionary-based word replacement or definition insertion to approach the need. However, these methods are limited by expert curation, which is hard to scale and has trouble generalizing to unseen datasets that do not share an overlapping vocabulary. In contrast, we approach the clinical word and sentence translation problem in a completely unsupervised manner. We show that a framework using representation learning, bilingual dictionary induction and statistical machine translation yields the best precision at 10 of 0.827 on professional-to-consumer word translation, and mean opinion scores of 4.10 and 4.28 out of 5 for clinical correctness and layperson readability, respectively, on sentence translation. Our fully-unsupervised strategy overcomes the curation problem, and the clinically meaningful evaluation reduces biases from inappropriate evaluators, which are critical in clinical machine learning.
BibTeX
@inproceedings{weng2019unsupervised,
  author = {Wei-Hung Weng and Yu-An Chung and Peter Szolovits},
  title = {{Unsupervised Clinical Language Translation}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2019},
  doi = {10.1145/3292500.3330710},
}
Figure