Year: 2026
Type: conference-paper
Abstract: Extracting structured clinical information by selecting a text span from long medical discharge summaries using Natural Language Processing (NLP) poses difficult challenges.Heterogeneity of the text, semantic overlap of the decision categories, and uncertainty in identifying boundaries of text span aligned with highly unbalanced labels.Contribute to these challenges, the MedExACT competition in BioNLP @ ACL 2026 formulates the medical concept and text span extraction problem from unstructured text as a classification task over nine DICTUM decision categories.Unstructured ICU discharge summaries are used from the MedDec dataset, for solution approach across eight demographic subgroups.We propose a two-stage data processing approach for the text span extraction and classification task where Stage 1 performs document-level sequence labeling with a pipeline composed of a RoBERTabase, a Bidirectional Gated Recurrent Unit (Bi-GRU), and a Conditional Random Field (CRF).Stage 2 applies a local span refinement module which revisits each candidate span to adjust both category assignments and boundary offsets.Our framework achieves a final score of 0.5668 which substantially outperforms the provided baseline.Ablation results further suggest that BiGRU-based contextual consolidation improves CRF-based decoding in text span extraction and classification.
Authors: Jing Tao, Amir Eskandari, Farhana Zulkernine
Institution: Queen's University
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: gold
OpenAlex ID: W7166809116