Counterfactual Reasoning for Fair Clinical Risk Prediction

Abstract

The use of machine learning systems to support decision making in healthcare raises questions as to what extent these systems may introduce or exacerbate disparities in care for historically underrepresented and mistreated groups, due to biases implicitly embedded in observational data in electronic health records. To address this problem in the context of clinical risk prediction models, we develop an augmented counterfactual fairness criteria to extend the group fairness criteria of equalized odds to an individual level. We do so by requiring that the same prediction be made for a patient, and a counterfactual patient resulting from changing a sensitive attribute, if the factual and counterfactual outcomes do not differ. We investigate the extent to which the augmented counterfactual fairness criteria may be applied to develop fair models for prolonged inpatient length of stay and mortality with observational electronic health records data. As the fairness criteria is illdefined without knowledge of the data generating process, we use a variational autoencoder to perform counterfactual inference in the context of an assumed causal graph. While our technique provides a means to trade off maintenance of fairness with reduction in predictive performance in the context of a learned generative model, further work is needed to assess the generality of this approach.

Failure Modes of Variational Autoencoders and Their Effects on Downstream Tasks

“…However, as VAEs are increasingly being used in application where the data is numeric, e.g. in medical or financial domains [37,19,50], these intuitive qualitative checks no longer apply. For example, in many medical applications, the original data features themselves (e.g.…”

Adversarial Learning for Counterfactual Fairness

“…It leverages the previous work [24], which introduced a causal framework to learn from biased data by exploring the relationship between sensitive features and data. With the recent development in deep learning, some novels approaches [25,26,27] argue to lead to a less error-prone decision-making model, by improving the approximation of the causal inference in the presence of unobserved confounders.…”

“…It therefore compares the predictions of the same individual with an alternate version of him/herself. Similar extension can be done to adapt the Equalized Odds objective for the Counterfactual framework [26]. Learning transformations XA←a for a given causal graph is at the heart of Counterfactual Fairness, as described in the next subsection.…”

Counterfactual Fairness with Disentangled Causal Effect Variational Autoencoder

“…To compensate this potential problem in CEVAE, modified version of CEVAE (Pfohl et al 2019), or mCEVAE, assumed that x and y are caused by a and u. M mCEVAE uses the maximum mean discrepancy (MMD) to regularize the generations to remove the information of a from u, but this MMD regularization removes u components that is simply correlated to a, not caused by a. M mCEVAE = E q(u|a,x) λ x log p(x|a, u) + λ y log p(y|a, u)…”

“…For example, a causal model estimates the effect of sensitive variables, such as race and gender, on an admission result (Kusner et al 2017). Another study shows a causal model predicting a medication's effect on a patient's prognosis (Pfohl et al 2019). If we focus on modeling the exogenous uncertainty with Variational Autoencoder (VAE) (Louizos et al 2017;Pfohl et al 2019), it has been a common practice to introduce a single latent variable to reflect all exogenous uncertainty.…”