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

Area of Science:

• Medical Informatics
• Machine Learning
• Causal Inference

Background:

• Machine learning (ML) models in healthcare risk perpetuating existing biases.
• Existing fairness definitions may not fully capture complex healthcare contexts.

Purpose of the Study:

• To develop a generalizable pipeline for training causally fair ML models in observational healthcare settings.
• To address both direct and indirect sources of bias in ML models.
• To leverage foundation models for causally fair predictions.

Main Methods:

• Mapping structural fairness models to observational healthcare data.
• Developing a pipeline that incorporates healthcare context and disparities.
• Detangling direct and indirect bias sources to assess fairness-accuracy tradeoffs.
• Utilizing foundation models for downstream fair predictions.

Main Results:

• A model-agnostic pipeline for training causally fair ML models is presented.
• The pipeline effectively addresses both direct and indirect healthcare biases.
• Expanded characterizations of the fairness-accuracy tradeoff are provided.

Conclusions:

• The developed pipeline offers a method for creating fairer ML models in healthcare.
• This approach can mitigate biases stemming from clinical decisions and systemic disparities.
• Foundation models can be adapted for causally fair applications in healthcare.