Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Causal Framework for Mitigating Data Shifts in Healthcare

Kurt Butler, Stephanie Riley, Damian Machlanski, Edward Moroshko, Panagiotis Dimitrakopoulos, Thomas Melistas, Akchunya Chanchal, Konstantinos Vilouras, Zhihua Liu, Steven McDonagh, Hana Chockler, Ben Glocker, Niccolo Tempini, Matthew Sperrin, Sotirios A Tsaftaris, Ricardo Silva

Abstract

Developing predictive models that perform reliably across diverse patient populations and heterogeneous environments is a core aim of medical research. However, generalization is only possible if the learned model is robust to statistical differences between data used for training and data seen at the time and place of deployment. Domain generalization methods provide strategies to address data shifts, but each method comes with its own set of assumptions and trade-offs. To apply these methods in healthcare, we must understand how domain shifts arise, what assumptions we prefer to make, and what our design constraints are. This article proposes a causal framework for the design of predictive models to improve generalization. Causality provides a powerful language to characterize and understand diverse domain shifts, regardless of data modality. This allows us to pinpoint why models fail to generalize, leading to more principled strategies to prepare for and adapt to shifts. We recommend general mitigation strategies, discussing trade-offs and highlighting existing work. Our causality-based perspective offers a critical foundation for developing robust, interpretable, and clinically relevant AI solutions in healthcare, paving the way for reliable real-world deployment.

A Causal Framework for Mitigating Data Shifts in Healthcare

Abstract

Developing predictive models that perform reliably across diverse patient populations and heterogeneous environments is a core aim of medical research. However, generalization is only possible if the learned model is robust to statistical differences between data used for training and data seen at the time and place of deployment. Domain generalization methods provide strategies to address data shifts, but each method comes with its own set of assumptions and trade-offs. To apply these methods in healthcare, we must understand how domain shifts arise, what assumptions we prefer to make, and what our design constraints are. This article proposes a causal framework for the design of predictive models to improve generalization. Causality provides a powerful language to characterize and understand diverse domain shifts, regardless of data modality. This allows us to pinpoint why models fail to generalize, leading to more principled strategies to prepare for and adapt to shifts. We recommend general mitigation strategies, discussing trade-offs and highlighting existing work. Our causality-based perspective offers a critical foundation for developing robust, interpretable, and clinically relevant AI solutions in healthcare, paving the way for reliable real-world deployment.
Paper Structure (21 sections, 3 figures)

This paper contains 21 sections, 3 figures.

Table of Contents

  1. Main
  2. Sources of Variation in Healthcare and Medicine
  3. Population-level shifts
  4. Exposure, selection and observation-level shifts
  5. Definition-level shifts
  6. Understanding How Shifts Arise in Data
  7. A primer on augmented causal graph notation
  8. Classical domain generalization
  9. What causality can tell us about shifts
  10. Other forms of causal information
  11. Mapping Problems to Solutions
  12. I. Problem Conceptualization
  13. II. Causal Description
  14. III. Feasibility Analysis
  15. IV. Deployment Strategies
  16. ...and 6 more sections

Figures (3)

  • Figure 1: An illustration of how classical domain generalization differs from the causal theory of domain generalization. In this example, we show that even in a simple setting, a covariate shift could arise from several distinct causal scenarios.
  • Figure 2: An illustration of various ways in which causal descriptions enhance our understanding of data shifts and inform our strategies to address them. Causal descriptions, either as graphs or descriptions as a data-generation process, are useful to mitigate data shifts. As we later show in Fig. \ref{['fig:roadmap']}, formulating causal descriptions is a crucial part of our framework.
  • Figure 3: A roadmap of our framework for developing predictive models. We define four stages: I. Problem conceptualization, II. Causal description, III. Feasibility analysis, and IV. Deployment strategies. While the stages can be followed sequentially, model design is iterative and may require revisiting earlier stages as we evaluate the feasibility of the model.