A0443
Title: Algorithmic fairness of models for predicting Alzheimer's disease progression
Authors: Kristin Linn - University of Pennsylvania (United States) [presenting]
Abstract: Alzheimer's disease (AD) disproportionately affects marginalized older adults. Machine learning (ML) techniques have the potential to improve early detection of AD. However, ML models may suffer from biases and perpetuate existing disparities. The fairness of three ML models for predicting progression from normal cognition to mild cognitive impairment (MCI) and from MCI to AD is audited. Three common fairness metrics are assessed (equal opportunity, equalized odds, and demographic parity), measured across subgroups defined by gender, ethnicity, and race. Although the three models demonstrated high accuracy in aggregate, all three models failed to satisfy fairness metrics for subgroups defined by ethnicity and race. The models generally satisfied metrics of fairness for gender. Potential implications of the findings are discussed and placed in context with recently published literature on algorithmic fairness.
