CASPAR: A Context-Aware Span Refinement Approach for Decision Support

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research, Data Science & Analytics · Depth: Expert, quick

Summary

CASPAR, a two-stage approach, addresses the MedExACT shared task for medical decision span extraction and classification from ICU discharge summaries. The first stage employs a sliding-window RoBERTa encoder with BiGRU and CRF for document-level sequence labeling, generating initial candidate spans. Subsequently, a lightweight refinement module in Stage 2 re-examines each candidate within its surrounding context to adjust category assignments and rectify span boundaries. This system achieved a final score of 0.5668 on the official leaderboard, significantly surpassing the organizer baseline in span-level F1. The paper also includes ablation results, repeated-run validation statistics, and detailed subgroup- and error-level analyses, which underscore difficulties in precise boundary recovery and confusion within race categories subgroups during clinical decision extraction.

Key takeaway

For NLP Engineers developing clinical text extraction systems, CASPAR's two-stage approach offers a robust strategy to enhance accuracy. You should consider implementing a context-aware refinement module to correct initial span boundary and category errors, particularly when dealing with complex medical summaries. This method can significantly improve F1 scores, but be mindful of challenges in exact boundary recovery and potential biases in specific subgroups like race categories.

Key insights

CASPAR uses a two-stage context-aware span refinement to improve medical decision extraction from ICU summaries.

Principles

• Contextual refinement improves span boundary and category accuracy.
• Two-stage processing can address complex extraction tasks.

Method

Stage 1: Sliding-window RoBERTa with BiGRU and CRF for candidate span generation. Stage 2: Lightweight module refines candidate spans and categories using surrounding context.

In practice

• Apply two-stage span refinement for complex NLP tasks.
• Use context-aware modules to correct initial span errors.
• Evaluate boundary recovery and subgroup performance.

Topics

• Medical Decision Extraction
• Clinical NLP
• Span Refinement
• RoBERTa
• Sequence Labeling
• ICU Discharge Summaries

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Predicting relations between SOAP note sections: The value of incorporating a clinical information model — Fine-tuning a RoBERTa-base model on manually annotated clinical data improves NER for SOAP note sections.
• BCL: Bayesian In-Context Learning Framework for Information Extraction — BCL uses Bayesian particle filtering to systematically optimize in-context learning for information extraction, improving performance and generalizability.
• Spancat: a new approach for span labeling — The SpanCategorizer (Spancat) is a spaCy component for structured annotation of diverse, complex, and overlapping text spans.
• Explicit Evidence Grounding via Structured Inline Citation Generation — LLMs excel at document retrieval but struggle with precise evidence span identification for faithful inline citations.
• Severity-Aware Curriculum Learning with Multi-Model Response Selection for Medical Text Generation — Severity-aware curriculum learning with multi-model response selection significantly improves medical text generation quality and relevance.
• Retrieval-Augmented Generation for Clinical Question Answering in Portuguese Drug Leaflets: Benefits and Limitations — RAG improves factual recall in clinical Q&A but needs adaptive control for complex reasoning to avoid anchoring bias.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.