Equitable Early Prediction of ICU Length of Stay from First-Day Clinical Data

Early prediction of Intensive Care Unit (ICU) length of stay (LOS) supports resource planning and clinical decision-making, yet standard evaluations often overlook disparities across patient subgroups. This study reframes early ICU LOS prediction as a fairness-critical task, examining how model design influences subgroup equity. Using first-day data from the MIMIC-III database, we train machine learning models for binary long-stay prediction and optimise decision thresholds via Pareto-based precision-recall trade-offs. We assess performance by gender and ethnicity and apply fairness mitigation methods such as threshold optimisation and exponentiated gradient post-processing. Models with strong overall accuracy show marked subgroup disparities under a single global threshold. Fairness-aware thresholding and post-processing substantially reduce these gaps while maintaining predictive performance, underscoring decision threshold selection as a key lever for equitable and transparent ICU prediction systems.