FairGATDANN: A Fairness-Aware Graph Attention and Domain-Adversarial Neural Network for ICU Mortality Prediction

This thesis introduces FairGATDANN, a fairness-aware machine learning model designed for ICU mortality prediction. FairGATDANN integrates Graph Attention Networks (GAT), domain-adversarial neural networks (DANN), and adaptive fairness constraints to address key challenges in healthcare AI, including predictive accuracy, fairness, explainability, and robustness to domain shifts. The model incorporates GraphSMOTE to address class imbalance, a Gradient-Reversal Layer to achieve domain-invariant representations, and utilizes dynamic dual-ascent optimization to maintain adaptive fairness constraints (here, Equalized Odds). Evaluations using the MIMIC-IV, MIMICIII, and eICU datasets demonstrate that FairGATDANN has the potential to achieve predictive performance comparable or superior to traditional ensemble methods, such as the Random Forest classifier or XGBoost, and reduces fairness disparities. Thus, FairGATDANN presents as a robust and compliant solution in line with regulatory frameworks such as the EU AI Act, contributing significantly to the development of equitable and reliable healthcare AI solutions.