Title: FairGATE: a fairness aware machine learning package

Abstract: The application of machine learning in clinical prediction is often hindered by algorithmic bias, which can perpetuate health disparities for marginalised demographic groups. In-processing fairness techniques, which integrate bias mitigation directly into model training, offer a promising solution but often lack interpretability. This study introduces and evaluates a novel, interpretable Gated Neural Network (GNN) FairGATE (https://cran.r-project.org/web/packages/fairGATE/index.html) designed to produce equitable predictions by explicitly modelling subgroup heterogeneity. A GNN with a custom fairness-constrained loss function was developed in R using the torch package. The model was trained to predict different outcomes while minimising disparities in error rates across protected attributes. Its performance was benchmarked against a standard XGBoost model using the IBM AI Fairness 360 toolkit in Python. Interpretability analyses, including gate weight and feature importance assessments, were conducted to understand the GNN's internal mechanisms. The GNN achieved a statistically significant improvement in Disparate Impact, corroborated by significant improvements in Statistical Parity Difference confirming that the model reduced systematic bias against some protected features. Crucially, the GNN produced better-calibrated probability predictions for unprivileged subgroups, correcting the overconfidence bias of the baseline model. We conclude that the GNN architecture is a viable and interpretable method for mitigating demographic bias in clinical prediction tasks. By navigating the accuracy-fairness trade-off, it produces more equitable and reliable predictions, offering a tangible step towards the development of more trustworthy and responsible clinical AI. 

Speaker Biography: 

Rhys Holland graduated in Biomedical Sciences and holds an MSc in Applied Statistical modelling and Health Informatics from King’s College London. He is currently working as a Clinical Data Analyst at PharosAI within Breast Cancer Research. 

Join with this Teams meeting link.

Please contact maudsley.brc@kcl.ac.uk if you have any questions.