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

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

Introduction

In modern healthcare, medical reports serve as the primary communication channel between specialists. However, a significant "expectation gap" exists: while doctors write these reports for other professionals, focusing on specific pathological findings, patients are increasingly accessing these documents. For a non-expert, the primary goal is often to answer a single, critical question: "Is there anything abnormal?"

Unfortunately, medical reports are frequently opaque to patients due to three main factors: medical jargon (e.g., "unremarkable" meaning normal), contradictory findings (e.g., "low normal"), and misleading grammatical structures that obscure the diagnostic intent.

Examples of ambiguity in medical reports and their disambiguated versions.

While previous Natural Language Processing (NLP) efforts have focused on "laymanization"—the simple replacement of complex terms—this research introduces a framework for disambiguation. The goal is to ensure the clinical intent and diagnostic certainty are clear to non-experts without compromising medical accuracy.

Methodology: Contrastive Knowledge Infusion

The authors propose a two-stage framework for Medical Report Disambiguation, designed to maintain high clinical fidelity while improving clarity.

The proposed framework consisting of (a) Contrastive Pretraining and (b) the Rewriting Framework.

1. Contrastive Pretraining

The process begins with a BART (Seq2Seq) model pretrained on a large-scale medical corpus (MIMIC-CXR). To ensure the model captures clinical nuances, the researchers employ Supervised Contrastive Learning. This technique forces the model to cluster sentences with similar underlying pathologies closely in its latent space. By doing so, the model learns that "unremarkable" and "normal" are functionally equivalent in a medical context, even if they are linguistically distinct.

2. Perturbation-Based Rewriting

The second stage involves a "plug-and-play" rewriting mechanism. The system first identifies "ambiguous tokens" using a specialized classifier. During the generation phase, the model's hidden states are "steered" or perturbed. This "nudges" the AI to produce outputs that are more decisively "Normal" or "Abnormal," aligning with the original report's intent while maximizing clarity. This approach allows for "minimal edits," preserving the core clinical description while resolving confusing phrasing.

Experimental Setup and Datasets

The framework was evaluated using three primary data sources:

  • OpenI-Annotated: Chest X-ray reports annotated by radiology specialists for ambiguity and abnormality.
  • VA-Annotated: A multi-modal dataset covering 8 modalities and 35 body parts, also professionally annotated.
  • MIMIC-CXR: A massive public dataset of ~1.1 million chest X-ray sentences used for the initial contrastive pretraining.

Key Results

The proposed model, Med-DEPEN, was benchmarked against several baselines, including Knowledge-Based Replacement (KBR) and standard Style Transformers.

Quantitative and Qualitative Performance

Med-DEPEN demonstrated a superior balance between Disambiguation (reduction in ambiguity) and Pathology Match (preservation of the original diagnosis). Traditional style-transfer models often failed by "destroying" clinical content in the pursuit of simplicity. In contrast, Med-DEPEN maintained high fidelity through its contrastive latent space.

In human evaluations, radiology experts found that Med-DEPEN achieved a 79.2% success rate in disambiguation on the OpenI dataset while maintaining 92.1% fidelity to the original content. A notable success was the transformation of the confusing "The lung volumes are low normal" into the clearer "The lung volumes are in the lower half of the normal limit."

Conclusion and Future Work

This work addresses the critical need for patient-centered medical communication. By infusing clinical knowledge into the latent space of generative models, the Med-DEPEN framework successfully disambiguates complex reports without sacrificing accuracy. Future work will focus on integrating expert auditing workflows and expanding the system to handle more diverse medical specialties beyond radiology.

Summary updated

Abstract

Sharing medical reports is essential for patient-centered care. A recent line of work has focused on automatically generating reports with NLP methods. However, different audiences have different purposes when writing/reading medical reports -- for example, healthcare professionals care more about pathology, whereas patients are more concerned with the diagnosis ("Is there any abnormality?"). The expectation gap results in a common situation where patients find their medical reports to be ambiguous and therefore unsure about the next steps. In this work, we explore the audience expectation gap in healthcare and summarize common ambiguities that lead patients to be confused about their diagnosis into three categories: medical jargon, contradictory findings, and misleading grammatical errors. Based on our analysis, we define a disambiguation rewriting task to regenerate an input to be unambiguous while preserving information about the original content. We further propose a rewriting algorithm based on contrastive pretraining and perturbation-based rewriting. In addition, we create two datasets, OpenI-Annotated based on chest reports and VA-Annotated based on general medical reports, with available binary labels for ambiguity and abnormality presence annotated by radiology specialists. Experimental results on these datasets show that our proposed algorithm effectively rewrites input sentences in a less ambiguous way with high content fidelity. Our code and annotated data are released to facilitate future research.
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