CUET_DiagNLP at #SMM4H-HeaRD 2026: Per-Axis TNM Staging from Pathology Reports and Opioid Impact Span Detection from Social Media

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Health & Wellbeing — Public Health & Epidemiology, Clinical Care & Medical Practice · Depth: Advanced, medium

Summary

CUET_DiagNLP developed systems for two #SMM4H-HeaRD 2026 shared tasks, focusing on medical text analysis and social media mining. For Task 6, which involved per-axis TNM cancer staging from free-text TCGA pathology reports, they fine-tuned GatorTron-base on each axis. This approach, utilizing Focal loss with class weights and a pooled [CLS]–mean representation, achieved macro F1 scores of 0.700 (T), 0.774 (N), and 0.640 (M) on test set 2, significantly outperforming the baseline scores of 0.454, 0.591, and 0.554. For Task 7, detecting span-level opioid-related ClinicalImpacts and SocialImpacts in Reddit posts, the team employed a uniform-weight ensemble combining DeBERTa-large and PubMedBERT. This system, enhanced with boundary-aware loss, entity-replacement augmentation, and a first-person post filter, yielded strict F1 of 0.51 and relaxed F1 of 0.60, surpassing both the task mean (0.46 / 0.55) and median (0.48 / 0.58).

Key takeaway

For NLP Engineers developing health information extraction systems, consider specialized fine-tuning and ensemble approaches. Your team can improve TNM staging accuracy from pathology reports by fine-tuning models like GatorTron-base per-axis, especially with imbalanced data. For social media health monitoring, combining models such as DeBERTa-large and PubMedBERT with augmentation and specific filters can significantly boost span detection F1 scores.

Key insights

Fine-tuning large language models and ensemble methods significantly improve medical text and social media health data extraction.

Principles

• Fine-tuning models per-axis improves specific task performance.
• Ensemble methods enhance F1 scores in span detection.
• Focal loss addresses label imbalance in classification.

Method

For TNM staging, fine-tune GatorTron-base per-axis with Focal loss and [CLS]–mean representation. For opioid impact, ensemble DeBERTa-large and PubMedBERT with boundary-aware loss, entity-replacement, and a first-person filter.

In practice

• Apply GatorTron-base for structured medical report extraction.
• Use DeBERTa/PubMedBERT ensembles for social media health insights.
• Implement data augmentation for improved span detection.

Topics

• TNM Staging
• Opioid Impact Detection
• Medical NLP
• Social Media Mining
• Large Language Models
• Ensemble Learning

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Domain Fine-Tuning FinBERT on Finnish Histopathological Reports: Train-Time Signals and Downstream Correlations — Domain fine-tuning FinBERT on medical text can be predicted by observing embedding geometry changes.
• Improved tumor-only variant calling and mutation burden estimation with VarNet-T — VarNet-T improves tumor-only somatic variant calling and TMB estimation using weakly supervised deep learning.
• 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.
• Automated Detection of Dosing Errors in Clinical Trial Narratives: A Multi-Modal Feature Engineering Approach with LightGBM — Combining sparse lexical features with dense semantic embeddings improves clinical dosing error detection.
• How to Fine-Tune an SLM for Emotion Recognition — Fine-tuning SLMs for fine-grained emotion recognition on imbalanced datasets requires a multi-technique approach to data balancing and loss weighting.
• Improving RCT-Based Treatment Effect Estimation Under Covariate Mismatch via Calibrated Alignment — CALM improves treatment effect estimation by aligning RCT and observational data in a shared embedding space.

Open in AIssential →

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