Thunderbolts at #SMM4H-HeaRD 2026: Detection of Insomnia in Clinical Notes using Transformers

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

Summary

The SuSh system, developed for Subtask 1 of the MultiClinAI shared task at the 11th SMM4H and HeaRD Workshop (ACL 2026), addresses multilingual clinical named entity recognition (NER) across seven languages. This system employs a fully zero-shot approach utilizing GLiNER-biomed-large-v1.0, a span-based NER model pre-trained on biomedical text, eliminating the need for task-specific fine-tuning or labeled data. To manage long clinical documents, it incorporates a character-level sliding window strategy. A post-processing pipeline further refines results through F1-max sweep for threshold optimization, entity-specific gazetteer lookup from DisTEMIST and SympTEMIST, span boundary correction, and negation filtering. The system achieved a Strict F1 of 0.5175, Strict Precision of 0.5536, Strict Recall of 0.4859, and CHR F1 of 0.6130 on the English disease subtask, demonstrating the competitiveness of domain-adapted zero-shot biomedical NER models as baselines.

Key takeaway

For NLP engineers developing clinical NER systems, you should consider zero-shot approaches with domain-adapted models like GLiNER-biomed-large-v1.0. This strategy can establish competitive baselines for multilingual tasks, significantly reducing the need for extensive labeled training data. Implement character-level sliding windows and a post-processing pipeline including F1-max optimization and gazetteer integration to maximize performance on long clinical notes.

Key insights

Domain-adapted zero-shot biomedical NER models offer competitive baselines for multilingual clinical entity recognition.

Principles

• Zero-shot NER can be competitive.
• Domain adaptation is crucial for performance.
• Post-processing enhances raw model output.

Method

The SuSh system uses GLiNER-biomed-large-v1.0 with a character-level sliding window, followed by F1-max sweep, gazetteer lookup (DisTEMIST, SympTEMIST), span boundary correction, and negation filtering.

In practice

• Use GLiNER-biomed-large-v1.0 for clinical NER.
• Implement sliding windows for long texts.
• Integrate gazetteers for entity refinement.

Topics

• Clinical Named Entity Recognition
• Zero-shot Learning
• Transformers
• GLiNER-biomed-large-v1.0
• Multilingual NLP
• Biomedical NLP

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Zero-Shot NER with GliNER and spaCy — GliNER enables zero-shot named entity recognition for both generic and domain-specific labels without training data.
• KliniskVestBERT: BERT Model Specialised to Norwegian Clinical Texts — Domain-specific pre-training significantly enhances BERT models for Norwegian clinical NLP tasks.
• How to Leverage NER and Advanced NLP Techniques for Life Sciences — NER and advanced NLP are essential for structuring and interpreting the vast unstructured text data in Life Sciences.
• A Proactive EMR Assistant for Doctor-Patient Dialogue: Streaming ASR, Belief Stabilization, and Preliminary Controlled Evaluation — Proactive EMR assistants require streaming ASR, belief stabilization, and objectified retrieval for real-time consultation support.
• Detecting HIV-Related Stigma in Clinical Narratives Using Large Language Models — LLMs can effectively detect HIV-related stigma in clinical notes, with encoder models outperforming generative models in zero-shot settings.
• 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.

Open in AIssential →

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