When Your Only Tool Is A Hammer: Ethical Limitations of Algorithmic Fairness Solutions in Healthcare Machine Learning

It is no longer a hypothetical worry that artificial intelligence - more specifically, machine learning (ML) - can propagate the effects of pernicious bias in healthcare. To address these problems, some have proposed the development of 'algorithmic fairness' solutions. The primary goal of these solutions is to constrain the effect of pernicious bias with respect to a given outcome of interest as a function of one's protected identity (i.e., characteristics generally protected by civil or human rights legislation. The technical limitations of these solutions have been well-characterized. Ethically, the problematic implication - of developers, potentially, and end users - is that by virtue of algorithmic fairness solutions a model can be rendered 'objective' (i.e., free from the influence of pernicious bias). The ostensible neutrality of these solutions may unintentionally prompt new consequences for vulnerable groups by obscuring downstream problems due to the persistence of real-world bias. The main epistemic limitation of algorithmic fairness is that it assumes the relationship between the extent of bias's impact on a given health outcome and one's protected identity is mathematically quantifiable. The reality is that social and structural factors confluence in complex and unknown ways to produce health inequalities. Some of these are biologic in nature, and differences like these are directly relevant to predicting a health event and should be incorporated into the model's design. Others are reflective of prejudice, lack of access to healthcare, or implicit bias. Sometimes, there may be a combination. With respect to any specific task, it is difficult to untangle the complex relationships between potentially influential factors and which ones are 'fair' and which are not to inform their inclusion or mitigation in the model's design.