GoBlueInformatics at #SMM4H-HeaRD 2026: Long-Context Encoders and Generative Biomedical LLMs for Pathological TNM Stage Prediction
Year: 2026
Type: conference-paper
Abstract: We describe our systems for #SMM4H-HeaRD 2026 Task 6, which requires predicting the T, N, and M components of pathological TNM stage from TCGA pathology reports.We explored both discriminative long-context encoders and generative biomedical LLMs.For the first test set, our BioClinical-ModernBERT-large ensemble achieved 0.993 micro-F1 and 0.915 macro-F1, improving over the BB-TEN baseline scoring-log result of 0.947 micro-F1 and 0.780 macro-F1.For the harder second test set, our OpenBioLLM-8B LoRA extractor improved component macro-F1 over the organizer baseline from 0.454 to 0.626 for T, from 0.591 to 0.758 for N, and from 0.554 to 1.000 for M.These results suggest that long-context encoders are strong for explicit T and N evidence, while constrained generative LLM extraction can be effective for harder reports.The main remaining weakness is rare-class T4 recognition.
Author: Shangqing Wei
Institution: University of Michigan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166807755