Identifying Semantics in Clinical Reports Using Neural Machine Translation

Srikanth Mujjiga, Vamsi Krishna, Kalyan Chakravarthi, Vijayananda JAAAI 2019 · DOI

Abstract

Medical concept extraction—the identification of standardized medical terms within clinical notes—is a critical component of modern healthcare informatics. However, traditional rule-based systems are often too slow for real-time applications. Researchers from Philips Healthcare propose a novel approach by framing concept mapping as a Neural Machine Translation (NMT) task. By "translating" English clinical sentences into Unified Medical Language System (UMLS) codes, their system achieves a 100x speedup over industry-standard tools while maintaining high retrieval accuracy.

Introduction

Radiologists and clinicians frequently need to navigate millions of historical reports to identify similar cases. Standard keyword-based Information Retrieval (IR) systems often fail to capture the semantic nuances of clinical language. For instance, a search for "heart attack" may fail to retrieve reports mentioning "myocardial infarction," despite their semantic equivalence. Furthermore, handling negations (e.g., "no evidence of pneumonia") is essential to avoid false positives in search results.

While tools like MetaMap are designed to map text to the Unified Medical Language System (UMLS), they are computationally intensive, often requiring several seconds per document. This latency is prohibitive for large-scale, real-time production environments. To address this, the authors developed a high-speed semantic search pipeline.

System architecture for clinical text processing and search query mapping.

Methodology

The primary innovation of this work is the conceptualization of medical concept mapping as a translation problem.

Framing the Translation Task

In this framework, the source language is standard English clinical text, and the target "language" is a sequence of "Enriched UMLS concepts." To handle negation within the translation, the authors appended a suffix to the UMLS Concept Unique Identifier (CUI). For example, a concept is represented as C0027051_0 if present and C0027051_1 if negated.

Model Architecture

The researchers evaluated several encoder-decoder architectures to optimize performance. The most effective model utilized a hybrid CNN+RNN encoder paired with an attention mechanism. The CNN component excels at capturing local n-gram and phrase-level features, while the RNN (specifically an LSTM) manages long-term dependencies and the overall sequence structure of the clinical narrative.

Hybrid CNN+RNN encoder-decoder architecture with attention mechanism.

Embeddings and Transfer Learning

To better capture medical context, the authors eschewed generic word embeddings in favor of transfer learning. They utilized the skip-gram model on a large corpus of clinical data (TREC-CDS) to refine pre-trained embeddings, ensuring the model understood the specific vocabulary of the medical domain.

Datasets and Evaluation

The model was trained on the TREC-CDS (Clinical Decision Support) dataset, which contains approximately 5 million sentences. Due to the scale of the data, a "silver standard" training set was generated using MetaMap in conjunction with a negation detector. For final evaluation, the system was tested against 5,000 clinical reports across multiple modalities (X-ray, MRI, CT, etc.) using 50 expert-validated queries.

Results

The NMT-based approach demonstrated significant improvements in both efficiency and accuracy:

  • Efficiency: The system reduced processing time from 3,000 milliseconds to just 30 milliseconds per document—a 100x speedup.
  • Translation Quality: The hybrid CNN+RNN model achieved a BLEU score of 40.1 and a perplexity of 3.03.
  • Search Performance: In retrieval tests, the system achieved a Mean Average Precision (MAP@10) of 0.76.

Conclusion and Future Work

By treating medical coding as a translation task, the authors have enabled context-aware semantic search that is viable for real-time clinical use. Future work will focus on improving the model's handling of long-distance negations and expanding the target vocabulary to include more niche medical terms. Additionally, the researchers plan to incorporate user feedback loops to further refine search rankings and customization.

Summary updated

BibTeX
@inproceedings{mujjiga2019identifying,
  author = {Srikanth Mujjiga and Vamsi Krishna and Kalyan Chakravarthi and Vijayananda J},
  title = {{Identifying Semantics in Clinical Reports Using Neural Machine Translation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33019552},
}
Figure