Introduction
Extracting structured information from clinical text is a complex task. Unlike general-domain text, medical narratives contain long, nested mentions with ambiguous boundaries. For instance, determining whether "massive heart attack" is a single trigger or a combination of a descriptor and a trigger requires high precision. Researchers from UCLA have introduced DICE (Data-effIcient Clinical Event extraction), a generative framework designed to address these challenges, particularly when annotated data is scarce.
The Challenges of Clinical Event Extraction
Event Extraction (EE) typically involves identifying "triggers" (e.g., "biopsy") and their associated "arguments" (e.g., "liver", "2023-10-01"). Clinical EE is uniquely difficult due to:
- Data Scarcity: HIPAA regulations and the need for expert annotators limit dataset availability.
- Vague Boundaries: Clinical mentions are often 50% longer than those in general news, frequently including complex descriptors.
- Argument Density: Clinical events often involve a high number of unique argument roles, creating a "long-tail" distribution where rare roles are difficult to learn.
Methodology
DICE leverages the T5-large generative architecture, treating extraction as a sequence-to-sequence task. It introduces three key innovations to handle medical nuances:
1. Contrastive Learning with Perturbation
To solve the boundary problem, DICE uses contrastive learning. It generates "negative" examples by slightly shifting mention boundaries (e.g., removing "the" or "right" from "the right common carotid artery"). The model is trained to distinguish the exact correct span from these perturbations.
2. Auxiliary Mention Identification (MI)
DICE incorporates a secondary task to identify any meaningful clinical mention. This "warm-up" task helps the model internalize the general structure of medical language before attempting specific event categorization.
3. Independent Role Querying
Rather than extracting all arguments in one pass, DICE queries for each role independently (e.g., "What is the Biological_structure?"). This approach prevents rare roles from being overshadowed by more common ones in the output sequence.
Experimental Results
The researchers evaluated DICE on MACCROBAT-EE, a new clinical dataset with full argument annotations, and the general-domain ACE05 dataset.
Performance and Data Efficiency
DICE achieved state-of-the-art (SOTA) results, outperforming strong baselines like OneIE by 2.7 F1 points in clinical argument classification. Its most significant advantage is observed in low-resource settings. When trained on only 10% of the available data, DICE outperformed the Text2Event baseline by over 8 F1 points, proving its robustness when expert annotations are limited.
Conclusion
DICE demonstrates that generative models, when enhanced with mention-aware techniques and contrastive learning, can overcome the inherent difficulties of clinical text. While the independent querying approach increases computational cost, the gains in accuracy and data efficiency make it a powerful tool for specialized domains where high-quality data is at a premium.
Summary updated