Fair Machine Learning in Healthcare: A Review

TL;DR

The paper addresses fairness in healthcare ML by applying distributive justice to classify issues into equal allocation and equal performance, and by mapping these principles to concrete fairness metrics. It provides a comprehensive lifecycle view of biases—from data collection through deployment—and reviews a broad set of mitigation strategies, correlating them with the identified bias sources. The work highlights the need for careful evaluation of mitigation methods against healthcare-specific fairness metrics and discusses challenges like uncertainty, long-term effects, multimodal data, and ethics. Collectively, it offers a structured framework to align ethical and technical dimensions of fair ML in healthcare and suggests directions to advance trustworthy, equitable AI in medical practice.

Abstract

The digitization of healthcare data coupled with advances in computational capabilities has propelled the adoption of machine learning (ML) in healthcare. However, these methods can perpetuate or even exacerbate existing disparities, leading to fairness concerns such as the unequal distribution of resources and diagnostic inaccuracies among different demographic groups. Addressing these fairness problem is paramount to prevent further entrenchment of social injustices. In this survey, we analyze the intersection of fairness in machine learning and healthcare disparities. We adopt a framework based on the principles of distributive justice to categorize fairness concerns into two distinct classes: equal allocation and equal performance. We provide a critical review of the associated fairness metrics from a machine learning standpoint and examine biases and mitigation strategies across the stages of the ML lifecycle, discussing the relationship between biases and their countermeasures. The paper concludes with a discussion on the pressing challenges that remain unaddressed in ensuring fairness in healthcare ML, and proposes several new research directions that hold promise for developing ethical and equitable ML applications in healthcare.

Figures (4)

• Figure 1: Bias at the different stages in machine learning systems: Red and blue represent two demographic groups. (a) The biases that exist at the data collection stage include minority bias, missing-data bias and label bias. Minority bias occurs when the sample size of the demographic groups are unbalanced. Missing data bias occurs when data may be missing in a non-random way. Label bias occurs when the quality of labels varies between different demographic groups. (b) Algorithm bias exists in model development stage, leads to systematical unfair results for certain demographic group. (c) The biases that exist at the data collection stage include interaction bias and training-serving skew bias. Training-serving skew bias occurs when the distribution of data in the deployment stage differs from the distribution of data in the training phase. Interaction bias occurs patients and healthcare professionals interact with machine learning models. Please refer to section \ref{['sec:source']} for further details.
• Figure 2: An illustration of methods mitigating the fairness problem in data collection stage: (a) Data redistribution methods adjust the distribution of the data. The diversified collection method collects data from other hospitals. The reweighting method assigns the weights to minority data. The resampling method seeks to create fair training samples in the sampling strategy. The synthetic method generates fake data. (b) Data purification methods remove sensitive information directly from the data. For example, removing sensitive attributes from tabular data or removing gender-specific pronouns from textual data.
• Figure 3: An illustration of methods mitigating the fairness problem in model development stage: (a) Model desensitization removes the ability of the model to discriminate between sensitive attribute information. Adversarial learning disables the model of predicting sensitive attributes. Disentanglement method separates and removes the sensitive attribute information from latent embedding. Contrastive learning enforces the samples with various sensitive attributes to be close in latent space. (b) Model constraint methods add additional constraints or regularization term.
• Figure 4: An illustration of methods mitigating the fairness problem in model deployment stage: (a) The decision explanation method offers the explanation to the outcome via XAI tool. (b) The model adjustment method fine-tunes the last few layers of the deployed mode. (c) The outcome adjustment method adjusts the original outcome to meet fairness requirement.