Predicting Angiographic Disease Status: Drawing the line between demographically decoupled and jointly trained models

Objective Application of machine learning predictive models for determining angiographic disease status have the risk of amplifying existing biases towards demographic groups based on age and gender. We formalize the underlying choice between demographically decoupled and jointly trained models and propose a framework that allows domain experts to balance equity with the goal of achieving the most accurate classifier for all demographic groups. Materials and Methods We propose an objective called Demographic Pareto Efficiency to discover classifiers for diagnosing angiographic disease status that optimize the demographic group accuracies for four groups based on age (>60, <=60) and gender (male, female). We discover predictive models on the Pareto frontier of group-level accuracy and aid domain experts in making efficient accuracy trade-offs. Results We outperform baseline classifiers incorporating min-max, adversarial and parity based notions of fairness both in overall accuracy and group-level accuracy by up to 9.7% and 9.6% respectively while retaining comparable levels of discrepancy between groups in the UCI Heart Disease dataset. Our approach searches for Pareto optimal group performance, whereas baseline approaches converge to non-pareto solutions, thus leaving room for improvement in group level accuracies. Discussion In determining angiographic disease status, machine learning predictive models need to maximize the accuracy of stratified demographic groups based on age and gender, by leveraging the benefits of transfer learning across groups through iterative and group-aware joint training, rather than maximizing an overall demographic group-agnostic accuracy measure. Conclusion Demographic Pareto Efficiency provides a framework to maximize prediction accuracy across demographic groups, while retaining fairness within a relaxation bound.