Fair predictions in health

ended

In clinical care, machine learning is progressively used to enhance diagnosis, therapy choice, and effectiveness of the health system. Because machine-learning models learn from historically gathered information, populations that have suffered past human and structural biases (e.g. unequal access to education or resources) — called protected groups — are susceptible to damage from inaccurate projections or resource allocations, reinforcing health inequalities. For example, racial and gender differences exist in the way clinical data are produced and these can be transferred as biases in the models. Several techniques of algorithmic fairness have been suggested in the literature on machine learning to ameliorate the performance of machine learning with respect to its fairness. The debate in statistics and machine learning has however failed to provide a principled approach for choosing concepts of bias, prejudice, discrimination, and fairness in predictive models, with a clear link to ethical theory discussed within philosophy. The specific scientific objectives of this research project are: O1: ethical theory: mapping the ethical theories that are relevant for the allocation of resources in health care and draw connections with the literature in fair machine learning O2: probabilistic ethics: understand how standard moral concepts such as responsibility, merit, need, talent, equality, and benefit can be understood in probabilistic terms O3: epistemology of causality: understand if current claims made by counterfactual and causal models of fairness in AI are robust with respect to different philosophical understandings of probability, causality, and counterfactuals O4: application: to show the relevance these philosophical ideas by applying them to a limited number of paradigmatic cases of the application of predictive algorithms in health care.