Algorithmic Fairness of Machine Learning Models for Alzheimer's Disease Progression Prediction

IntroductionAlzheimers disease (AD) disproportionately affects older adults from marginalized communities. Predictive models using machine learning (ML) techniques have potential to improve early detection and management of AD. However, ML models potentially suffer from biases and may perpetuate or exacerbate existing disparities. MethodsWe investigated algorithmic fairness of logistic regression, support vector machines and recurrent neural networks for predicting progression to mild cognitive impairment and AD. Fairness was quantified across gender, ethnicity, and race subgroups using three measures: equal opportunity, equalized odds and demographic parity. ResultsAll three ML models performed well in aggregate but demonstrated disparate performance across race and ethnicity subgroups. Compared to Non-Hispanic participants, sensitivity for predicting progression to mild cognitive impairment and to AD was 5%-9.6% and 16.8%-24.9% lower, respectively, for Hispanic participants. Sensitivity was similarly lower for Black and Asian participants compared to Non-Hispanic White participants. Models generally satisfied metrics of fairness with respect to gender. DiscussionAlthough accurate in aggregate, models failed to satisfy fairness metrics. Fairness should be considered in the development and deployment of ML models for AD progression.