A3S@C-DAC at #SMM4H-HeaRD 2026: Reasoning Meets Evidence: LLMs for Interpretable Insomnia Detection with Evidence Extraction in Clinical Notes

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Health & Wellbeing — Health & Medical Research, Clinical Care & Medical Practice, Medical Devices & Health Technology · Depth: Expert, quick

Summary

A3S@C-DAC's systems for the SMM4H-HeaRD 2026 shared task address insomnia detection from clinical narratives, focusing on accurate classification and interpretable evidence extraction. The team utilized MIMIC-III notes, which were annotated with rule-based insomnia criteria and supporting evidence spans. Two primary approaches were explored: parameter-efficient fine-tuning (PEFT) of lightweight models using QLoRA and LoRA, and few-shot prompting of large language models (LLMs) for combined reasoning and evidence extraction. Their top-performing system achieved an F1-score of 0.7333 for binary classification and a micro-F1 of 0.6535 for multi-label rule prediction. Evidence extraction reached up to 0.5192 partial-match F1. The findings indicate that lightweight fine-tuned models excel in classification, while larger models demonstrate superior reasoning capabilities but struggle with precise evidence span localization, highlighting a critical need in clinically interpretable NLP.

Key takeaway

For NLP Engineers developing interpretable clinical systems, you should consider a hybrid approach. Fine-tune lightweight models using PEFT techniques like QLoRA or LoRA for robust classification, as they demonstrated superior performance (F1-score of 0.7333). Simultaneously, integrate larger language models for their stronger reasoning capabilities, but be prepared to address their current limitations in precise evidence span localization (up to 0.5192 partial-match F1). This strategy can balance classification accuracy with the need for clinically grounded, interpretable evidence.

Key insights

Lightweight fine-tuned models classify insomnia better, while LLMs reason stronger but struggle with precise evidence span localization.

Principles

• Interpretable clinical NLP requires both classification and evidence.
• PEFT can enable lightweight models to outperform larger LLMs.
• LLMs excel at reasoning but need better span localization.

Method

Systems used parameter-efficient fine-tuning (QLoRA, LoRA) on lightweight models and few-shot prompting of LLMs for joint reasoning and evidence extraction from MIMIC-III notes.

In practice

• Fine-tune lightweight models for classification tasks.
• Use LLMs for complex reasoning in clinical notes.
• Annotate clinical notes with rule-based criteria and evidence.

Topics

• Insomnia Detection
• Clinical NLP
• Large Language Models
• Parameter-Efficient Fine-Tuning
• Evidence Extraction
• MIMIC-III

Best for: AI Scientist, Research Scientist, NLP Engineer

Related on AIssential

• Active Evidence-Seeking and Diagnostic Reasoning in Large Language Models for Clinical Decision Support — LLMs exhibit a substantial performance gap in active, multi-turn clinical diagnosis compared to static, full-context evaluations.
• LLM Fine-tuning: Techniques for Adapting Language Models — Fine-tuning LLMs with PEFT methods like LoRA and QLoRA significantly reduces computational demands while preserving performance.
• Small Updates, Big Doubts: Does Parameter-Efficient Fine-tuning Enhance Hallucination Detection ? — PEFT enhances LLM hallucination detectability by reshaping uncertainty signals, not primarily by improving factual accuracy.
• Explainable Detection of Depression Status Shifts from User Digital Traces — A BERT-LLM framework detects mental health shifts from digital traces, providing explainable, temporally coherent reports.
• Small LLMs for Biomedical Claim Verification: Cost-Effective Fine-Tuning, Structural Dataset Shortcuts, and Cross-Domain Generalization — Small, QLoRA-fine-tuned LLMs can surpass larger proprietary models for specialized biomedical claim verification tasks.
• EASE-TTT: Evidence-Aligned Selective Test-Time Training for Long-Context Question Answering — Retrieved evidence can directly supervise query-side attention adaptation for improved long-context QA in smaller LLMs.

Open in AIssential →

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