Enhancing Fairness in Disease Prediction by Optimizing Multiple Domain Adversarial Networks

Bin Li, Xinghua Shi, Hongchang Gao, Xiaoqian Jiang, Kai Zhang, Arif Harmanci, Bradley MalinbioRxiv 2023 · DOI

Abstract

Predictive models in biomedicine need to ensure equitable and reliable outcomes for the populations they are applied to. Unfortunately, biases in medical predictions can lead to unfair treatment and widening disparities, underscoring the need for effective techniques to address these issues. To enhance fairness, we introduce a framework based on a Multiple Domain Adversarial Neural Network (MDANN), which incorporates multiple adversarial components. In an MDANN, an adversarial module is applied to learn a fair pattern by negative gradients back-propagating across multiple sensitive features. We leverage loss functions based on the Area Under the Receiver Operating Characteristic Curve (AUC) to address the class imbalance, promoting equitable classification performance for minority groups. Moreover, we utilize pre-trained convolutional autoencoders (CAEs) to extract deep representations of data, aiming to enhance prediction accuracy and fairness. Combining these mechanisms, we alleviate biases and disparities to provide reliable and equitable disease prediction. We empirically demonstrate that the MDANN approach leads to better accuracy and fairness in predicting disease progression using brain imaging data for Alzheimer's Disease and Autism populations than state-of-the-art techniques.

Introduction

Artificial Intelligence in medicine holds great promise for "precision medicine," but it faces significant challenges regarding bias. If a model is trained on data where certain groups—based on race, gender, or age—are underrepresented, it often learns to favor the majority, leading to "one-size-fits-all" outcomes that fail vulnerable populations. Medical datasets are notoriously imbalanced, and real-world bias is often intersectional, involving multiple sensitive features simultaneously. Most existing fairness techniques struggle to mitigate these multiple biases without compromising the model's predictive power.

The MDANN Framework

The proposed Multiple Domain Adversarial Neural Network (MDANN) framework addresses these issues through three synergistic pillars: deep feature extraction, adversarial learning with gradient reversal, and a specialized loss function for imbalanced data.

The MDANN architecture: A pre-trained CAE extracts features, which are then processed by a predictor and an adversarial module. The gradient reversal layer ensures the feature extractor "unlearns" sensitive demographic information.

Methodology Components

  1. Feature Extractor (Pre-trained CAE): Instead of raw images, the model uses a Convolutional Autoencoder (CAE) to compress brain images into high-dimensional representations. This architecture captures essential textures while filtering out noise that often carries demographic bias.
  2. Adversarial Module with Gradient Reversal: The model includes multiple Adversarial Components (ACs) tasked with predicting sensitive features like sex or age. During training, a gradient reversal layer flips the gradients, forcing the feature extractor to learn representations that make it impossible for these sub-modules to succeed, effectively "unlearning" discriminatory information.
  3. AUC-induced Minimax Loss: Standard cross-entropy loss often fails with imbalanced data. MDANN implements a minimax loss function based on the Area Under the Curve (AUC), forcing the optimizer to prioritize the classification performance of minority groups.

Experimental Evaluation

The framework was evaluated on two major biomedical repositories: the ABIDE II (Autism) dataset, targeting sex and handedness, and the ADNI (Alzheimer’s) dataset, targeting age, education, and handedness.

Key Results

  • Predictive Performance: In Alzheimer’s prediction (ADNI), the model achieved a peak validation accuracy of 0.85, significantly higher than the 0.76 baseline.
  • Fairness Metrics: Using Disparate Impact (DI) as a metric (where 1.0 is perfect fairness), MDANN achieved a DI of 0.841 for sex in Autism prediction, compared to just 0.386 for previous state-of-the-art methods.
  • Handling Imbalance: In the highly imbalanced handedness category, the AUC-minimax loss boosted minority group accuracy from 0.6 to 0.8, demonstrating the model's ability to maintain performance for underrepresented groups.

Conclusion

The MDANN approach demonstrates that fairness and accuracy are not necessarily at odds. By utilizing robust feature extractors and adversarial components to create domain-invariant features, the model provides more equitable disease prediction. Future work will explore tailoring adversarial architectures to specific sensitive features and incorporating generative models to capture even more nuanced biological patterns.

Summary updated

Abstract

Predictive models in biomedicine need to ensure equitable and reliable outcomes for the populations they are applied to. Unfortunately, biases in medical predictions can lead to unfair treatment and widening disparities, underscoring the need for effective techniques to address these issues. To enhance fairness, we introduce a framework based on a Multiple Domain Adversarial Neural Network (MDANN), which incorporates multiple adversarial components. In an MDANN, an adversarial module is applied to learn a fair pattern by negative gradients back-propagating across multiple sensitive features (i.e., characteristics of individuals that should not be used to discriminate unfairly between individuals when making predictions or decisions.) We leverage loss functions based on the Area Under the Receiver Operating Characteristic Curve (AUC) to address the class imbalance, promoting equitable classification performance for minority groups (e.g., a subset of the population that is underrepresented or disadvantaged.) Moreover, we utilize pre-trained convolutional autoencoders (CAEs) to extract deep representations of data, aiming to enhance prediction accuracy and fairness. Combining these mechanisms, we alleviate biases and disparities to provide reliable and equitable disease prediction. We empirically demonstrate that the MDANN approach leads to better accuracy and fairness in predicting disease progression using brain imaging data for Alzheimer's Disease and Autism populations than state-of-the-art techniques.
BibTeX
@misc{li2023enhancing,
  author = {Bin Li and Xinghua Shi and Hongchang Gao and Xiaoqian Jiang and Kai Zhang and Arif Harmanci and Bradley Malin},
  title = {{Enhancing Fairness in Disease Prediction by Optimizing Multiple Domain Adversarial Networks}},
  year = {2023},
  doi = {10.1101/2023.08.04.551906},
  howpublished = {bioRxiv},
}
Figure