A pipeline for enabling path-specific causal fairness in observational health data

TL;DR

This work tackles bias in healthcare ML by adopting path-specific causal fairness within a structural fairness model (SFM) that segregates sensitive, mediator, confounder, and outcome variables. It introduces a general, model-agnostic pipeline to map observational health data to $X$, $Y$, $W$, $Z$, and to robustly estimate path-specific effects such as the natural direct effect ($NDE$), natural indirect effect ($NIE$), and spurious effect ($SE$); the pipeline then trains a baseline model and applies targeted fairness interventions (causal regularization, resampling, and feature selection) to mitigate disparities across four clinical tasks (AMI, SLE, T2DM, SCZ). The experiments show task-dependent tradeoffs between fairness and predictive performance, with outcomes varying by which causal path is prioritized and which intervention is used; naive feature selection sometimes matches or outperforms sophisticated causal methods in reducing path-specific bias. The work emphasizes context-specific prioritization over a universal fairness solution and demonstrates the approach’s generalizability to healthcare and potentially beyond. Overall, the pipeline provides a principled framework to diagnose, enforce, and evaluate path-specific fairness in observational health data, aligning model behavior with known health disparities while preserving clinical utility.

Abstract

When training machine learning (ML) models for potential deployment in a healthcare setting, it is essential to ensure that they do not replicate or exacerbate existing healthcare biases. Although many definitions of fairness exist, we focus on path-specific causal fairness, which allows us to better consider the social and medical contexts in which biases occur (e.g., direct discrimination by a clinician or model versus bias due to differential access to the healthcare system) and to characterize how these biases may appear in learned models. In this work, we map the structural fairness model to the observational healthcare setting and create a generalizable pipeline for training causally fair models. The pipeline explicitly considers specific healthcare context and disparities to define a target "fair" model. Our work fills two major gaps: first, we expand on characterizations of the "fairness-accuracy" tradeoff by detangling direct and indirect sources of bias and jointly presenting these fairness considerations alongside considerations of accuracy in the context of broadly known biases. Second, we demonstrate how a foundation model trained without fairness constraints on observational health data can be leveraged to generate causally fair downstream predictions in tasks with known social and medical disparities. This work presents a model-agnostic pipeline for training causally fair machine learning models that address both direct and indirect forms of healthcare bias.

Figures (14)

• Figure 1: Standard Fairness Model: (a) Overview of SFM as presented in plecko_cfatoolkit_2024 (b) Clinically motivated mapping of SFM for observational health data
• Figure 2: Overview of the proposed workflow. Starting with task selection, we outline the steps required to train and evaluate a causally fair model. In particular, we find that data-driven decision making must be combined with known historical context about healthcare disparities specific to the clinical task at hand.
• Figure 3: Error in the path-specific causal estimates for dimensionality $|W|=1,750$. Each line corresponds to a different dimensionality reduction method, with the black line corresponding to no dimensionality reduction. The NDE estimation error is reflected in part (a), and the NIE estimation error is reflected in part (b). We find that interventions that identify features important for predicting the outcome ($Y$) outperform all other methods.
• Figure 4: Tradeoffs between model performance and target path-specific effects across all tasks. We plot the target path-specific effects for each intervention (NDE, NIE, and/or SE) against the model performance (AUROC) for each of the "target" effects (Section \ref{['sec:methods_selectingeffects']}). (A) and (E) show the NDE and NIE for AMI; (B) shows the NDE for SLE; (C) and (F) show the NDE and NIE for T2DM; (D) and (G) show the NDE and SE for SCZ.
• Figure 5: Error in the path-specific causal estimates for the (a) natural direct effect (NDE), (b) natural indirect effect (NIE), and (c) spurious effect (SE). Error is highest for a small sample size. The estimates are generated using doubly robust estimation, but still demonstrate large error for NDE and NIE.