CheXclusion: Fairness gaps in deep chest X-ray classifiers

By Laleh Seyyed-Kalantari, Guanxiong Liu, Matthew B. A. McDermott, Marzyeh Ghassemi. Pacific Symposium on Biocomputing, 2020

385 citations recorded by the source corpus.

Machine learning systems have received much attention recently for their ability to achieve expert-level performance on clinical tasks, particularly in medical imaging. Here, we examine the extent to which state-of-the-art deep learning classifiers trained to yield diagnostic labels from X-ray images are biased with respect to protected attributes. We train convolution neural networks to predict 14 diagnostic labels in 3 prominent public chest X-ray datasets: MIMIC-CXR, Chest-Xray8, CheXpert, as well as a multi-site aggregation of all those datasets. We evaluate the TPR disparity - the difference in true positive rates (TPR) - among different protected attributes such as patient sex, age, race, and insurance type as a proxy for socioeconomic status. We demonstrate that TPR disparities exist in the state-of-the-art classifiers in all datasets, for all clinical tasks, and all subgroups. A multi-source dataset corresponds to the smallest disparities, suggesting one way to reduce bias. We find that TPR disparities are not significantly correlated with a subgroup's proportional disease burden. As clinical models move from papers to products, we encourage clinical decision makers to carefully audit for algorithmic disparities prior to deployment. Our supplementary materials can be found at, http://www.marzyehghassemi.com/chexclusion-supp-3/.

Subfields: Health Informatics, Radiology, Nuclear Medicine and Imaging

View Paper: https://www.semanticscholar.org/paper/54c6416eae67bce7b3f058601e0c758e39f33a62

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Health Informatics

In Radiology, Nuclear Medicine and Imaging

Citations in the corpus

Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations (2021, methodology): “However, they are evaluated in each of five models (the same model trained five times with five different random seeds 7 ), with the train– validation–test split kept fixed in the training of the five models.”
Hurtful words: quantifying biases in clinical contextual word embeddings (2020, background): “In clinical applications, machine learning has the potential to improve patient outcomes, cut costs, and reduce physician burnout [22, 58, 63].”
Characteristics of publicly available skin cancer image datasets: a systematic review. (2021, background): “26,29,30 35 This underlines the importance of detailing the exact composition of datasets through metadata reporting, to ensure the generalisability of algorithms to real-world populations.”
Shifting machine learning for healthcare from development to deployment and from models to data (2022, background): “Training models on multi-institutional datasets can be most effective at combating model deterioration 15 , and directly combating existing biases in the training data can also mitigate their impact 171 .”
COVID-19 Image Data Collection: Prospective Predictions Are the Future (2020, background): “This can lead to many negative outcomes if treated as a typical benchmark dataset [Maguolo and Nanni, 2020; Cohen et al., 2020b; Seyyed-Kalantari et al., 2020; Kelly et al., 2019].”
AUC Maximization in the Era of Big Data and AI: A Survey (2022, background): “Although these issues have received tremendous attention in the literature for medical image classification [26, 145, 146], developing fair and interpretable DAMmethods remains to be explored.”
Generative Adversarial Networks (GANs) (2020, background): “[229] identified that existing deep learning classifiers trained to generate diagnostic labels from X-ray images are biased w.”
Ethical Machine Learning in Healthcare (2020, background): “Choosing these patients can rely on parts of the EHR that may be skewed due to lack of access to care, or abnormalities in clinical care: e.g., economic incentives may alter diagnosis code logging (81), clinical protocol affects the frequency and observation of abnormal tests (62), historical racial”
AI recognition of patient race in medical imaging: a modelling study (2021, neutral): “10 In related work, several groups reported that AI algorithms can identify various demographic patient factors.”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “…have found relatively small disparities in algorithmic performance across demographic groups (Chen et al., 2019; Larrazabal et al., 2020); Seyyed-Kalantari et al. (2020) finds larger disparities in true positive rates across demographic groups, and future work should investigate whether…”
RadGraph: Extracting Clinical Entities and Relations from Radiology Reports (2021, methodology): “[29], when deploying clinical models in practice, researchers should audit performance disparities across attributes, such as sex, age, and race, which we report in Table 2 for our datasets.”