@inproceedings{wei-2026-goblueinformatics,
    title = "{G}o{B}lue{I}nformatics at {\#}{SMM}4{H}-{H}ea{RD} 2026: Long-Context Encoders and Generative Biomedical {LLM}s for Pathological {TNM} Stage Prediction",
    author = "Wei, Shangqing",
    editor = "Lopez-Garcia, Guillermo  and
      Gonzalez-Hernandez, Graciela",
    booktitle = "Proceedings of the 11th Social Media Mining for Health Research and Applications ({SMM}4{H}-{H}ea{RD} 2026) Workshop and Shared Tasks",
    month = jul,
    year = "2026",
    address = "San Diego, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.smm4h-1.18/",
    pages = "108--112",
    ISBN = "979-8-89176-432-3",
    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."
}