Year: 2026
Type: conference-paper
Abstract: We build an ensemble of 10 transformer encoders for the MedExACT 2026 shared task on medical decision span detection.The ensemble is diversified along three training directions: encoder initialization (including domain-adaptive pre-training on clinical text), loss function, and data augmentation with LLM-generated synthetic notes and silver-labeled clinical documents.Greedy forward search selects the combination with the highest validation score.The system achieves a final score of 58.09 (Span F1 51.81, Token F1 66.66) on the test set.Analysis shows that each added model differs from existing members along at least one training direction, producing span predictions that allow majority voting to screen out individual errors.
Institutions: Siriraj Hospital, Mahidol University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7166871147