SubmissionNumber#=%=#30 FinalPaperTitle#=%=#Sparse Category Routing and Fairness-Aware Optimization for Medical Decision Extraction ShortPaperTitle#=%=# NumberOfPages#=%=# CopyrightSigned#=%=# JobTitle#==# Organization#==# Abstract#==#Extracting structured medical decisions from ICU discharge summaries is hard because of long documents, severe category imbalance across nine DICTUM decision types, and a fairness-aware evaluation that penalizes incon- sistent performance across demographic sub- groups. We present our system for the MedEx- ACT 2026 shared task (Elgaar et al., 2026), which fine-tunes BiomedBERT with a com- posite loss combining label-smoothed cross- entropy, a soft token-F1 auxiliary term, and R-Drop regularization. At inference time we apply a deterministic ensemble: half-offset sliding-window augmentation across four win- dow configurations, dual-branch logit aggrega- tion from the same checkpoint, per-category length calibration on the Anchor Branch, and sparse routing of categories 4 and 7 to a context- weighted specialist branch motivated by their unusual span-length distributions. Adding R- Drop improved validation Overall_F1 by 1.24 points over the CE + soft-F1 baseline, with a larger 1.70-point gain on Worst-Group F1. Our best submission achieves Span F1 of 0.4900, Token F1 of 0.6796, and an official Overall_F1 of 0.5724, with the African American subgroup as the Worst-Group bottleneck at Base_Score 0.5601 Author{1}{Firstname}#=%=#Ahmed Author{1}{Lastname}#=%=#Elshehaby Author{1}{Orcid}#=%=# Author{1}{Email}#=%=#ahmed.elshehaby@mbzuai.ac.ae Author{1}{Affiliation}#=%=#UG student at MBZUAI Author{2}{Firstname}#=%=#Mohamed Author{2}{Lastname}#=%=#Abdalla Author{2}{Orcid}#=%=# Author{2}{Email}#=%=#mohamed.abdalla@mbzuai.ac.ae Author{2}{Affiliation}#=%=#UG student at MBZUAI Author{3}{Firstname}#=%=#Youssef MK Author{3}{Lastname}#=%=#Mohamed Author{3}{Username}#=%=#youssefkhalil320 Author{3}{Orcid}#=%=# Author{3}{Email}#=%=#youssef.mohamed@mbzuai.ac.ae Author{3}{Affiliation}#=%=#MS student at MBZUAI ========== èéáğö