Abstract
Unstructured clinical texts contain rich health-related information. To better utilize the knowledge buried in clinical texts, discovering synonyms for a medical query term has become an important task. Recent automatic synonym discovery methods leveraging raw text information have been developed. However, to preserve patient privacy and security, it is usually quite difficult to get access to large-scale raw clinical texts. In this paper, we study a new setting named synonym discovery on privacy-aware clinical data (i.e., medical terms extracted from the clinical texts and their aggregated co-occurrence counts, without raw clinical texts). To solve the problem, we propose a new framework SurfCon that leverages two important types of information in the privacy-aware clinical data, i.e., the surface form information, and the global context information for synonym discovery. In particular, the surface form module enables us to detect synonyms that look similar while the global context module plays a complementary role to discover synonyms that are semantically similar but in different surface forms, and both allow us to deal with the OOV query issue. We conduct extensive experiments and case studies on publicly available privacy-aware clinical data, and show that SurfCon can outperform strong baseline methods by large margins under various settings.
Introduction
Unstructured clinical notes are a goldmine for medical knowledge, but they are also a privacy minefield. Due to HIPAA and other regulations, researchers rarely get access to raw patient records. Instead, institutions often release "privacy-aware" datasets: lists of extracted medical terms and their aggregated co-occurrence counts (how often two terms appear near each other). Traditional synonym discovery relies heavily on "local context"—the specific sentences where a word appears. Without raw text, this context is lost.
Medical synonym discovery presents unique challenges. First, surface similarity can be misleading or helpful (e.g., "Vit C" vs. "Vitamin C"). Second, semantic similarity often exists without surface similarity (e.g., "Vitamin C" vs. "Ascorbic acid"). Finally, the "Out-of-Vocabulary" (OOV) problem occurs when users search for terms not present in the provided co-occurrence graph. SurfCon addresses these by combining surface form analysis with a dynamic context-matching mechanism.
The SurfCon Framework
SurfCon is designed to operate on privacy-aware data by leveraging two main modules: Bi-level Surface Form Encoding and Context Matching.
Bi-level Surface Form Encoding
This module analyzes the "look" of a term. It utilizes both character-level embeddings (to capture misspellings and abbreviations) and word-level embeddings. These are combined into a single vector representing the term's appearance, providing a strong prior for synonyms that share linguistic roots.
Context Matching and Prediction
To handle semantic similarity and OOV terms, SurfCon employs a two-step context process:
- Inductive Context Prediction: Raw co-occurrence graphs are often sparse or pruned for privacy. SurfCon uses the surface form vector to predict what a term's global context (its neighbors) should be. This allows the model to "hallucinate" a neighborhood for OOV terms that are not in the original graph.
- Dynamic Matching Mechanism: When comparing two terms, the model uses an attention-like mechanism to weight neighbors relevant to both terms. For example, when comparing "vitamin C" and "ascorbic acid," it prioritizes shared context like "iron absorption" or "anemia."
Finally, these scores are integrated and optimized using the ListNet ranking algorithm to produce a final recommended list of synonyms.
Experimental Setup
The framework was evaluated using publicly available medical term co-occurrence data extracted from 20 million clinical notes from Stanford Hospitals and Clinics. Ground truth for synonyms was established using the Unified Medical Language System (UMLS), where different strings mapping to the same Concept Unique Identifier (CUI) are considered synonyms. A specific OOV test set of 2,000 terms not present in the initial co-occurrence graph was created to test the model's robustness.
Results and Discussion
SurfCon consistently outperformed strong baselines, including Word2vec and specialized synonym discovery models like DPE, across all settings.
- In-Vocabulary Performance: SurfCon achieved a Mean Average Precision (MAP) of 0.9176, significantly higher than the 0.8565 achieved by the best surface-only baseline.
- OOV Robustness: For terms never seen in the graph, SurfCon maintained a high MAP of 0.8301. In contrast, context-only models like Word2vec cannot function on OOV terms.
- Qualitative Insights: Case studies showed the model successfully identified "aphonia" as a synonym for "unable to vocalize," despite having zero character overlap, demonstrating the power of the context-matching module.
The researchers found that prediction acts as a form of denoising for pruned privacy-aware graphs. Furthermore, the dynamic matching mechanism provided a ~4% performance boost over static averaging, proving that context relevance is pair-dependent.
Conclusion
SurfCon demonstrates that high-quality synonym discovery is possible even under strict privacy constraints. By combining surface-level features with inductive context prediction, the framework overcomes the limitations of sparse co-occurrence data and OOV queries. Future work may involve integrating more complex encoders like Transformers and extending the framework to other specialized domains such as finance or law, where privacy-aware data constraints are similarly prevalent.
Summary updated