Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled Environments

This study introduces a novel end-to-end framework designed to optimize patient outcomes and operational efficiency in intensive care units (ICUs) through fairness-aware and accurate length of stay (LOS) predictions within an IoT-enabled environment. Motivated by the concept of “fairness through unawareness,” our proposed framework employs a demographic feature exclusion strategy, preventing access to potentially discriminatory information, and thus enforcing an inductive bias that redirects learning toward non-discriminatory pattern dependencies. To address the loss of static information, we introduce a custom graph neural network that dynamically reconstructs patient relationships over time, adapting from static demographics to evolving inter-patient correlations via multi-modal embeddings (e.g., medications, procedures, vitals, conditions) and learned feature-driven edge formation. We evaluate our approach on MIMIC-IV across diverse time-series formats and demonstrate its superiority over top performing LOS prediction methods, showcasing notable performance improvements while maintaining comparable computational complexity. Additionally, we present a performance analysis highlighting our approach’s computational and predictive scalability across different graph sizes at inference, while underscoring its intrinsic ability to improve performance when deprived of demographic features by dynamically readjusting its expressive power, unlike all evaluated methodologies in this study. Finally, we present a comparative study against the literature’s top-performing LOS prediction approaches that highlights our method’s predictive superiority, demonstrating significant performance gains in predicting LOS exceeding 3 and 7 days, with improvements of up to 13.1% AUC and 10.3% ARP, and 14.1% AUC and 42.0% ARP, respectively. The code is available at https://github.com/icsa-hua/mimic_fairness_aware_gnn