"Nothing Abnormal": Disambiguating Medical Reports via Contrastive Knowledge Infusion

Zexue He, An Yan, Amilcare Gentili, Julian J. McAuley, Chun-Nan Hsu

Abstract

Medical reports are primarily authored for professional-to-professional communication, often resulting in an "expectation gap" when accessed by patients. Technical jargon, shorthand, and ambiguous phrasing can lead to patient anxiety or misunderstanding. This paper presents a novel Natural Language Processing (NLP) framework designed to rewrite ambiguous medical reports into clear, patient-friendly language. By employing a two-stage Contrastive Knowledge Infusion approach, the system—termed Med-DEPEN—explicitly clarifies diagnostic outcomes (e.g., distinguishing between normal and abnormal findings) while strictly preserving the underlying medical facts. Experimental results on specialized benchmarks demonstrate that this method significantly outperforms standard neural baselines in both disambiguation success and clinical fidelity.

Introduction

In the modern healthcare landscape, patients increasingly have direct access to their clinical records. However, the language used in these reports is optimized for clinical precision rather than lay comprehension. A doctor might describe a finding as "unremarkable," a term that sounds dismissive or even negative to a patient, but signifies a healthy, normal result in a medical context.

This communication gap is characterized by three primary types of ambiguity:

  1. Medical Jargon: Terms like "prominent" or "unremarkable" carry specific clinical weights that contradict their colloquial meanings.
  2. Contradictory Findings: Phrases such as "low normal" create uncertainty regarding whether a result is actionable.
  3. Grammatical and Syntactic Errors: The "telegraphic" style of medical writing—often missing punctuation or subjects—can make it unclear if a condition is being confirmed or ruled out.

Existing AI solutions for text simplification often struggle with these nuances, either over-simplifying to the point of losing critical detail or "hallucinating" new medical conditions that were not present in the original text.

Methodology: Contrastive Knowledge Infusion

To address these challenges, we propose the Med-DEPEN framework, which utilizes a two-stage process to perform "disambiguation rewriting."

1. Contrastive Pretraining

The foundation of the model is a Seq2Seq architecture (BART) pretrained on large-scale medical datasets, including MIMIC-CXR. The core innovation at this stage is the use of Supervised Contrastive Learning. This technique forces the model to group sentences with similar medical pathologies together in its latent representation space. By anchoring the model’s understanding to specific diagnostic categories, we ensure that any subsequent rewriting remains within the "neighborhood" of the original diagnosis, preventing the model from accidentally altering the clinical meaning (e.g., confusing a pulmonary issue with a cardiac one).

2. The Rewriting Framework: Detect and Perturb

Rather than performing a wholesale rewrite of the text—which increases the risk of error—Med-DEPEN uses a surgical "plug-and-play" approach:

  • Detect: An ambiguity classifier identifies specific tokens or phrases likely to cause patient confusion.
  • Perturb: The model then "nudges" its internal representations toward a clear diagnostic direction (Normal vs. Abnormal) using a dedicated abnormality classifier. This results in minimal, targeted edits that clarify the outcome without disturbing the surrounding medical facts.

Experimental Setup and Datasets

The framework was evaluated using several key datasets, including two newly developed benchmarks:

  • OpenI-Annotated: Chest radiology reports annotated by specialists for both ambiguity and abnormality.
  • VA-Annotated: A diverse dataset covering eight medical modalities and 35 body parts, providing a broad test of the model's generalizability.
  • MIMIC-CXR: Approximately 1.1 million chest X-ray report sentences used for the initial contrastive pretraining phase.

Results and Discussion

Med-DEPEN was compared against several baselines, including Style Transformers and Knowledge-Based Replacement (KBR) systems.

Human and Expert Evaluation

In evaluations conducted by professional radiologists, Med-DEPEN achieved a 0.792 disambiguation success rate on the OpenI dataset. This represents a substantial improvement over the next best neural baseline, which scored 0.609.

Content Fidelity

A critical metric in medical NLP is "pathology preservation." Med-DEPEN maintained a pathology match score of 0.921, indicating that the model successfully clarified the text without changing the underlying medical diagnosis. This high fidelity is attributed to the contrastive learning "anchor" that prevents the model from drifting into incorrect medical territory during the generation process.

The "Unremarkable" Paradox

The study confirmed that minimal edits are often the most effective. By focusing on the most ambiguous tokens—such as mapping "unremarkable" to clear "normal" indicators—the model bridges the gap between clinical accuracy and patient understanding without the risks associated with long-form paraphrasing.

Conclusion and Future Work

While Med-DEPEN represents a significant step forward in patient-centered AI, the authors emphasize that such systems should currently serve as assistive tools rather than autonomous reporters. Future work will focus on multi-modal integration, allowing the AI to analyze medical images (such as X-rays or CT scans) alongside the text to provide even more robust disambiguation. By refining the way medical information is communicated, we can reduce patient anxiety and foster a more transparent healthcare environment.

Summary updated

BibTeX
@inproceedings{he2023nothing,
  author = {Zexue He and An Yan and Amilcare Gentili and Julian J. McAuley and Chun-Nan Hsu},
  title = {{"Nothing Abnormal": Disambiguating Medical Reports via Contrastive Knowledge Infusion}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i12.26665},
}
Figure