Reducing Sample Selection Bias in Clinical Data through Generation of Multi-Objective Synthetic Data
Jarren Briscoe, Chance DeSmet, Katherine A. Wuestney, Assefaw Hadish Gebremedhin, Roschelle Lynette Fritz, Diane J. Cook · Proceedings of the World Congress on Electrical Engineering and Computer Systems and Science · 2024
In the era of data-driven healthcare, identifying, quantifying, and mitigating bias in machine learning is of paramount importance.The impact of fair machine learning is particularly significant when predictions are applied in a clinical setting, where biased predictions can lead to unequal healthcare outcomes.In this paper, we consider the area of biomedical informatics and examine existing bias metrics and introduce a new metric to analyze bias in a smart home dataset.We investigate bias that may occur along sensitive attributes and examine its impact on the machine learning task of activity recognition from the collected data.In a novel approach to bias mitigation, we introduce a multi-objective generative adversarial network that creates synthetic data to mitigate sample bias by enhancing data diversity.We validate these methods using data collected for older adults living in smart homes who are managing multiple chronic health conditions, highlighting the potential of our approach to improve health predictions and outcomes.