SubmissionNumber#=%=#15
FinalPaperTitle#=%=#Diverse Transformer Ensemble with Majority Voting for Medical Decision Extraction at MedExACT 2026
ShortPaperTitle#=%=#
NumberOfPages#=%=#
CopyrightSigned#=%=#
JobTitle#==#
Organization#==#
Abstract#==#We describe our submission to the MedExACT 2026 shared task on extracting and classifying medical decisions from ICU discharge summaries into nine DICTUM categories. We frame the task as BIO token classification and train 25 diverse transformer models spanning 13 distinct pretrained architectures, including Longformer, DeBERTa-v3, RoBERTa, BioBERT, Bio_ClinicalBERT, BiomedBERT, SciBERT, BlueBERT, PubMedBERT, ELECTRA, ALBERT, XLNet, and BERT. Each model is trained with category-aware oversampling, focal loss, and group-aware sampling weighted by demographic frequency to address class imbalance and promote fairness across patient subgroups. Long documents are handled via sliding-window chunking, and span boundaries are refined to match
the official evaluation preprocessing. At inference, predictions from all 25 models are aggregated via text-normalized majority voting: a span is retained only if at least six models agree on its category and  normalized text. Our best submission achieves a final score of 0.5554 on the test set, exceeding the organizer-provided RoBERTa-base baseline (0.5111) by 4.4 points absolute and ranking 12th out of 37 submitted systems. We find that (1) architectural diversity matters more than model count or sophisticated aggregation, (2) simple majority voting outperforms learned ensemble filters that overfit the small validation set, and (3) cross-validation is essential for reliable model selection on small clinical
  datasets.
Author{1}{Firstname}#=%=#Rishik K.
Author{1}{Lastname}#=%=#Kondadadi
Author{1}{Username}#=%=#krishik
Author{1}{Orcid}#=%=#
Author{1}{Email}#=%=#konda052@umn.edu
Author{1}{Affiliation}#=%=#Student

==========
èéáğö