Previous studies in medical imaging have shown disparate abilities of artificial intelligence (AI) to detect a person’s race, yet there is no known correlation for race on medical imaging that would be obvious to human experts when interpreting the images. We aimed to conduct a comprehensive evaluation of the ability of AI to recognise a patient’s racial identity from medical images.
Using private (Emory CXR, Emory Chest CT, Emory Cervical Spine, and Emory Mammogram) and public (MIMIC-CXR, CheXpert, National Lung Cancer Screening Trial, RSNA Pulmonary Embolism CT, and Digital Hand Atlas) datasets, we evaluated, first, performance quantification of deep learning models in detecting race from medical images, including the ability of these models to generalise to external environments and across multiple imaging modalities. Second, we assessed possible confounding of anatomic and phenotypic population features by assessing the ability of these hypothesised confounders to detect race in isolation using regression models, and by re-evaluating the deep learning models by testing them on datasets stratified by these hypothesised confounding variables. Last, by exploring the effect of image corruptions on model performance, we investigated the underlying mechanism by which AI models can recognise race.
In our study, we show that standard AI deep learning models can be trained to predict race from medical images with high performance across multiple imaging modalities, which was sustained under external validation conditions (x-ray imaging [area under the receiver operating characteristics curve (AUC) range 0·91–0·99], CT chest imaging [0·87–0·96], and mammography [0·81]). We also showed that this detection is not due to proxies or imaging-related surrogate covariates for race (eg, performance of possible confounders: body-mass index [AUC 0·55], disease distribution [0·61], and breast density [0·61]). Finally, we provide evidence to show that the ability of AI deep learning models persisted over all anatomical regions and frequency spectrums of the images, suggesting the efforts to control this behaviour when it is undesirable will be challenging and demand further study.
The results from our study emphasise that the ability of AI deep learning models to predict self-reported race is itself not the issue of importance. However, our finding that AI can accurately predict self-reported race, even from corrupted, cropped, and noised medical images, often when clinical experts cannot, creates an enormous risk for all model deployments in medical imaging.
National Institute of Biomedical Imaging and Bioengineering, MIDRC grant of National Institutes of Health, US National Science Foundation, National Library of Medicine of the National Institutes of Health, and Taiwan Ministry of Science and Technology
including in many health-care applications, such as detection of melanoma,
and algorithms that aid the prediction of health-care use,
in which the performance of AI is stratified by self-reported race on a variety of clinical tasks.
Several studies have shown disparities in the performance of medical AI systems across race. For example, Seyyed-Kalantari and colleagues showed that AI models produce significant differences in the accuracy of automated chest x-ray diagnosis across racial and other demographic groups, even when the models only had access to the chest x-ray itself.
Importantly, if used, such models would lead to more patients who are Black and female being incorrectly identified as healthy compared with patients who are White and male. Moreover, racial disparities are not simply due to under-representation of these patient groups in the training data, and there exists no statistically significant correlation between group membership and racial disparities.
found that an AI model could predict sex and distinguish between adult and paediatric patients from chest x-rays, while other studies
reported reasonable accuracy at predicting the chronological age of patients from various imaging studies. In ophthalmology, retinal images have been used to predict sex, age, and cardiac markers (eg, hypertension and smoking status).
These findings, which show that demographic factors that are strongly associated with disease outcomes (eg, age, sex, and racial identity), are also strongly associated with features of medical images and might induce bias in model results, mirroring what is known from over a century of clinical and epidemiological research on the importance of covariates and potential confounding.
Many published AI models have conceptually amounted to simple bivariate analyses (ie, image features and their ability to predict clinical outcomes). Although more recent AI models have begun to consider other risk factors that conceptually approach multivariate modelling, which is the mainstay of clinical and epidemiological research, key demographic covariates (eg, age, sex, and racial identity) have been largely ignored by most deep learning research in medicine.
Evidence before this study
We used three different search engines to do our review. For PubMed, we used the following search terms: “(((disparity OR bias OR fairness) AND (classification)) AND (x-ray OR mammography)) AND (machine learning [MeSH Terms]).” For IEEE Xplore, we used the following search terms: “((disparity OR bias OR fairness) AND (mammography OR x-ray) AND (machine learning))”. For ACM, we used the following search terms: “[Abstract: mammography x-ray] AND [Abstract: classification prediction] AND [All: disparity fairness]”. All queries were limited to dates between Jan 1, 2010, and Dec 31, 2020. We included any studies that were published in English, focused on medical images, and that were original research. We also reviewed commentaries and opinion articles. We excluded articles that were not written in English or that were outside of the medical imaging domain. To our knowledge, there is no published meta-analysis or systematic review on this topic. Most published papers focused on measuring disparities in tabular health data without much emphasis on imaging-based approaches.
Although previous work has shown the existence of racial disparities, the mechanism for these differences in medical imaging is, to the best of our knowledge, unexplored. Pierson and colleagues noted that an artificial intelligence (AI) model that was designed to predict severity of osteoarthritis using knee x-rays could not identify the race of the patients. Yi and colleagues conducted a forensics evaluation on chest x-rays and found that AI algorithms could predict sex, distinguish between adult and paediatric patients, and differentiate between US and Chinese patients. In ophthalmology, retinal scan images have been used to predict sex, age, and cardiac markers (eg, hypertension and smoking status). We found few published studies that explicitly targeted the recognition of racial identity from medical images, possibly because radiologists do not routinely have access to, nor rely on, demographic information (eg, race) for diagnostic tasks in clinical practice.
Added value of this study
In this study, we investigated a large number of publicly and privately available large-scale medical imaging datasets and found that self-reported race is accurately predictable by AI models trained with medical image pixel data alone as model inputs. First, we showed that AI models are able to predict race across multiple imaging modalities, various datasets, and diverse clinical tasks. This high level of performance persisted during external validation of these models across a range of academic centres and patient populations in the USA, as well as when the models were optimised to do clinically motivated tasks. Second, we conducted ablations that showed that this detection was not due to trivial proxies, such as body habitus, age, tissue density, or other potential imaging confounders for race (eg, underlying disease distribution in the population). Finally, we showed that the features learned appear to involve all regions of the image and frequency spectrum, suggesting the efforts to control this behaviour when it is undesirable will be challenging and demand further study.
Implications of all the available evidence
In our study, we emphasise that the ability of AI to predict racial identity is itself not the issue of importance, but rather that this capability is readily learned and therefore is likely to be present in many medical image analysis models, providing a direct vector for the reproduction or exacerbation of the racial disparities that already exist in medical practice. This risk is compounded by the fact that human experts cannot similarly identify racial identity from medical images, meaning that human oversight of AI models is of limited use to recognise and mitigate this problem. This issue creates an enormous risk for all model deployments in medical imaging: if an AI model relies on its ability to detect racial identity to make medical decisions, but in doing so produced race-specific errors, clinical radiologists (who do not typically have access to racial demographic information) would not be able to tell, thereby possibly leading to errors in health-care decision processes.
and, in contrast to other demographic factors (eg, age and sex), there is a widely held, but tacit, belief among radiologists that the identification of a patient’s race from medical images is almost impossible, and that most medical imaging tasks are essentially race agnostic (ie, the task is not affected by the patient’s race). Given the possibility for discriminatory harm in a key component of the medical system that is assumed to be race agnostic, understanding how race has a role in medical imaging models is of high importance
as many AI systems that use medical images as the primary inputs are being cleared by the US Food and Drug Administration and other regulatory agencies.
In this study, we aimed to investigate how AI systems are able to detect a patient’s race to differing degrees of accuracy across self-reported racial groups in medical imaging. To do so, we aimed to investigate large publicly and privately available medical imaging datasets to examine whether AI models are able to predict an individual’s race across multiple imaging modalities, various datasets, and diverse clinical tasks.
Definitions of race and racial identity
and are often incorrectly conflated with biological concepts (eg, genetic ancestry).
In this modelling study, we defined race as a social, political, and legal construct that relates to the interaction between external perceptions (ie, “how do others see me?”) and self-identification, and specifically make use of self-reported race of patients in all of our experiments. We variously use the terms race and racial identity to refer to this construct throughout this study.
Table 1Summary of datasets used for race prediction experiments
CXP=CheXpert dataset. DHA=Digital Hand Atlas. EM-CS=Emory Cervical Spine radiograph dataset. EM-CT=Emory Chest CT dataset. EM-Mammo=Emory Mammogram dataset. EMX=Emory chest x-ray dataset. MXR=MIMIC-CXR dataset. NLST=National Lung Cancer Screening Trial dataset. RSPECT=RSNA Pulmonary Embolism CT dataset.
Investigation of possible mechanisms of race detection
Table 2Summary of experiments conducted to investigate mechanisms of race detection in Black patients
BMI=body-mass index. CXP=CheXpert dataset. DHA=Digital Hand Atlas. EM-CS=Emory Cervical Spine radiograph dataset. EM-CT=Emory Chest CT dataset. EM-Mammo=Emory Mammogram dataset. EMX=Emory CXR dataset. MXR=MIMIC-CXR dataset. NLST=National Lung Cancer Screening Trial dataset. RSPECT=RSNA Pulmonary Embolism CT dataset.
We did not present measures of performance variance or null hypothesis tests because these data are uninformative given the large dataset sizes and the large effect sizes reported (ie, even in experiments in which a hypothesis could be defined, all p values were <0·001).
Race detection in radiology imaging
and Emory-chest x-ray (EMX) with both internal validation (ie, testing the model on an unseen subset of the dataset used to train the model) and external validation (ie, testing the model on a completely different dataset than the one used to train the model) to establish baseline performance. Second, we trained racial identity detection models for non-chest x-ray images from multiple body locations, including digital radiography, mammograms, lateral cervical spine radiographs, and chest CTs, to evaluate whether the model’s performance was limited to chest x-rays.
or breast density
). Second, we assessed whether there was a difference in disease distribution among patients of different racial groups (eg, previous studies provide evidence that Black patients have a higher incidence of particular diseases, such as cardiac disease, than White patients).
Third, we assessed whether there were location-specific or tissue-specific differences (eg, there is evidence that Black patients have a higher adjusted bone mineral density and a slower age-adjusted annual rate of decline in bone mineral density than White patients).
Fourth, we assessed whether there were effects of societal bias and environmental stress on race outcomes from medical imaging data, as shown by differences in race detection by age and sex (reflecting cumulative and occupational differences in exposures). Last, we assessed whether there was an effect on the ability of AI deep learning systems to detect race when multiple demographic and patient factors were combined, including age, sex, disease, and body habitus.
by evaluating, first, frequency domain differences in the high frequency image features (ie, textural) and low frequency image features (ie, structural) that could be predictive of race; second, how differences in image quality might influence the recognition of race in medical images (given the possibility that image acquisition practices might differ for patients with different racial identities); and, last, whether specific image regions contribute to the recognition of racial identity (eg, specific patches or regional variations in the images, such as radiographic markers in the top right corner).
Role of the funding source
Grant support was used to pay for data collection, data analysis, data interpretation, and writing of the manuscript. The funders did not influence the decision to publish or the target journal for publication.
Table 3Performance of deep learning models to detect race from chest x-rays
Values reflect the area under the receiver operating characteristics curve for each model on the test set per slice and per study (by averaging the predictions across all slices). CXP=CheXpert dataset. DHA=Digital Hand Atlas. EM-CS=Emory Cervical Spine radiograph dataset. EM-CT=Emory Chest CT dataset. EM-Mammo=Emory Mammogram dataset. EMX=Emory CXR dataset. MXR=MIMIC-CXR dataset. NLST=National Lung Cancer Screening Trial dataset. RSPECT=RSNA Pulmonary Embolism CT dataset.
We found that deep learning models effectively predicted patient race even when the bone density information was removed for both MXR (AUC value for Black patients: 0·960 [CI 0·958–0·963]) and CXP (AUC value for Black patients: 0·945 [CI 0·94–0·949]) datasets. The average pixel thresholds for different tissues did not produce any usable signal to detect race (AUC 0·5). These findings suggest that race information was not localised within the brightest pixels within the image (eg, in the bone).
In this modelling study, which used both private and public datasets, we found that deep learning models can accurately predict the self-reported race of patients from medical images alone. This finding is striking as this task is generally not understood to be possible for human experts. We also showed that the ability of deep models to predict race was generalised across different clinical environments, medical imaging modalities, and patient populations, suggesting that these models do not rely on local idiosyncratic differences in how imaging studies are conducted for patients with different racial identities. Beyond these findings, in two of the datasets (MXR and CXP) analysed, all patients were imaged in the same locations and with the same processes, presumably independently of race.
We also provide evidence that disease distribution and body habitus of patients in the CXP, MXR, and EMX datasets were not strongly predictive of racial group, implying that the deep learning models were not relying on these features alone. Although an aggregation of these and other features could be partially responsible for the ability of AI models to detect racial identity in medical images, we could not identify any specific image-based covariates that could explain the high recognition performance presented here.
which measured the extent to which models learned potentially sensitive attributes (eg, age, race, and BMI) from an institutional dataset (the AHRF dataset) of 1296 patient chest x-rays. Their findings led to an AUC value of 0·66 (0·54–0·79). Possible explanations for this discrepant performance compared with our experiment could be due to the use of transfer learning in Jabbour and colleagues’ study, in which the MXR and CXP datasets were used for initial training, and the final layers were fine-tuned on the AHRF dataset. This possible contamination in the dataset might have degraded performance due to label misalignment. We do not have access to the AHRF dataset for further external validation and Jabbour and colleagues did not extend their experiments to MXR and CXP datasets.
The results of the low-pass filter and high-pass filter experiments done in our study suggest that features relevant to the recognition of racial identity were present throughout the image frequency spectrum. Models trained on low-pass filtered images maintained high performance even for highly degraded images. More strikingly, models that were trained on high-pass filtered images maintained performance well beyond the point that the degraded images contained no recognisable structures; to the human coauthors and radiologists it was not clear that the image was an x-ray at all. Furthermore, experiments that were involved in patch-based training, slice-based error analysis, and saliency mapping were non-contributory: no specific regions of the images consistently informed race recognition decisions. Overall, we were unable to isolate specific image features that were responsible for the recognition of racial identity in medical images, either by spatial location, in the frequency domain, or that were caused by common anatomic and phenotype confounders associated with racial identity.
Although this approach has already been criticised as being ineffective, or even harmful in some circumstances,
our work suggests that such an approach could be impossible in medical imaging because racial identity information appears to be incredibly difficult to isolate. The ability to detect race was not mitigated by any reasonable reduction in resolution or by the addition of noise, nor by frequency spectrum filtering or patch-based masking. Even ignoring the question of whether these approaches were beneficial, it seems plausible that technical solutions along these lines are unlikely to succeed and that strategies designed to detect racial bias,
paired with the intentional design of models to equalise racial outcomes,
should be considered to be the default approach to optimise the safety and fairness of AI in this context. The regulatory environment in particular, while evolving, has not yet produced strong processes to guard against unexpected racial recognition by AI models; either to identify these capabilities in models or to mitigate the harms that might be caused.
We note that in the context of racial discrimination and bias, the vector of harm is not genetic ancestry but the social and cultural construct that of racial identity, which we have defined as the combination of external perceptions and self-identification of race. Indeed, biased decisions are not informed by genetic ancestry information, which is not directly available to medical decision makers in almost any plausible scenario. As such, self-reported race should be considered a strong proxy for racial identity.
Our study was also limited by the availability of racial identity labels and the small cohorts of patients from many racial identity categories. As such, we focused on Asian, Black, and White patients, and excluded patient populations that were too small to adequately analyse (eg, Native American patients). Additionally, Hispanic patient populations were also excluded because of variations in how this population was recorded across datasets. Moreover, our experiments to exclude bone density involved brightness clipping at 60% and evaluating average body tissue pixels, with no methods to evaluate if there was residual bone tissue that remained on the images. Future work could look at isolating different signals before image reconstruction.
The combination of reported disparities and the findings of this study suggest that the strong capacity of models to recognise race in medical images could lead to patient harm. In other words, AI models can not only predict the patients’ race from their medical images, but appear to make use of this capability to produce different health outcomes for members of different racial groups.
To conclude, our study showed that medical AI systems can easily learn to recognise self-reported racial identity from medical images, and that this capability is extremely difficult to isolate. We found that patient racial identity was readily learnable from medical imaging data alone, and could be generalised to external environments and across multiple imaging modalities. We strongly recommend that all developers, regulators, and users who are involved in medical image analysis consider the use of deep learning models with extreme caution as such information could be misused to perpetuate or even worsen the well documented racial disparities that exist in medical practice. Our findings indicate that future AI medical imaging work should emphasise explicit model performance audits on the basis of racial identity, sex, and age, and that medical imaging datasets should include the self-reported race of patients when possible to allow for further investigation and research into the human-hidden but model-decipherable information related to racial identity that these images appear to contain.
IB was responsible for the conceptualisation of the study, data curation from Emory, supervision of trainees, as well as writing, reviewing, and editing the manuscript. ARB was responsible for training the race prediction model for the Digital Hand Atlas, which was supervised by SP, as well as reviewing the manuscript and preparing the code repository accompanying the manuscript under the supervision of JWG. JLB participated in writing and reviewing the manuscript. LAC was responsible for the overall study design, critical review of the manuscript, as well as synthesis of the results, literature review, and writing and reviewing the manuscript. LC conducted the experiments on the MIMIC-CXR dataset under supervision of PK, reported results, and reviewed the manuscript. RC prepared the Emory chest x-ray dataset under supervision of JWG and IB, as well as contributing to the literature review and the review of the manuscript. ND conducted the experiments on anatomic and phenotypic confounders, specifically on predictions based on age and sex. MG was responsible for the overall study design, supervision of experiments, results analysis, manuscript writing, and literature review. JWG was responsible for the overall study design, Emory datasets extraction and curation, design and supervision of experiments, results analysis, literature review, and manuscript writing. JWG also provided qualitative review of the saliency maps to evaluate any localising information. SH created the Stanford RSPECT dataset under supervision of MPL and conducted external validation of the CT chest prediction model. PK was responsible for designing and conducting experiments on the MIMIC-CXR dataset for race prediction, exploration of anatomic and phenotypic confounders (including body-mass index [BMI]), segmenting the dataset into lung and non-lung segments, noisy and blurred images, ablation experiments, as well as writing and reviewing the manuscript. MPL was responsible for supervising the creation of the RSPECT Stanford dataset, extracting race labels for the CheXpert dataset, conducting qualitative review of saliency maps, as well as participating in the literature review and writing of the manuscript. BJP, supervised by LO-R and JWG, trained race prediction models on Emory cervical spine radiographs, Emory chest x-ray, MIMIC-CXR, and CheXpert datasets. BJP also conducted experiments on anatomic and phenotype confounders, specifically on the effect of resolution change on prediction and BMI. ATP assisted manuscript preparation and review. SP participated in the overall study and experiment design, supervised ARB, JLB, and BJP, and helped edit the manuscript. LO-R trained the CT chest race prediction model on the NLST dataset and supervised BJP on experiments. LO-R also designed the experiments and summarised results from multiple experiments, as well as manuscript writing and review. LO-R also conducted qualitative review of saliency maps. CO prepared the Emory chest x-ray dataset, under the supervision of JWG and IB, and conducted race prediction experiments on this dataset. LS-K assisted with the design of the overall study and experiments, critical review of results, literature review, and manuscript writing. HT prepared the Emory cervical spine radiographs and mammogram datasets with IB and JWG, and also contributed to writing and editing the manuscript. RW conducted the experiments on the MIMIC-CXR dataset, under the supervision of PK, reported results, and also reviewed the manuscript. ZZ, under the supervision of IB and JWG, prepared the Emory CT dataset, conducted the external validation of the CT experiments on the Emory dataset, trained race prediction models on the Emory mammogram dataset, summarised results, and reviewed the manuscript. HZ conducted the experiments on high and low filter image manipulations and reviewed the manuscript. LJP assisted with overall study design, critical review of results, and manuscript writing. All authors had access to the datasets used in this study. JWG, PK, IB, and HT verified the data.
Declaration of interests
MG has received speaker fees for a Harvard Medical School executive education class. HT has received consulting fees from Sirona medical, Arterys, and Biodata consortium. HT also owns lightbox AI, which provides expert annotation of medical images for radiology AI. MPL has received consulting fees from Bayer, Microsoft, Phillips, and Nines. MPL also owns stocks in Nines, SegMed, and Centaur. LAC has received support to attend meetings from MISTI Global Seed Funds. ATP has received payment for expert testimony from NCMIC insurance company. ATP also has a pending institutional patent for comorbidity prediction from radiology images. All other authors declare no competing interests.
JWG and ATP are funded by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) MIDRC grant of the National Institutes of Health (75N92020C00008 and 75N92020C00021). JWG and SP are funded by US National Science Foundation (grant number 1928481) from the Division of Electrical, Communication & Cyber Systems. MPL was funded by the National Library of Medicine of the National Institutes of Health (R01LM012966). LAC is funded by the National Institute of Health through a NIBIB grant (R01 EB017205). PK is funded by the Ministry of Science and Technology (Taiwan; MOST109-2222-E-007-004-MY3).
On the dangers of stochastic parrots: can language models be too big? FAccT ’21.
Racial disparities in automated speech recognition.
Proc Natl Acad Sci USA. 2020; 117: 7684-7689
Gender shades: intersectional accuracy disparities in commercial gender classification.
PMLR. 2018; 81: 77-91
Machine learning and health care disparities in dermatology.
JAMA Dermatol. 2018; 154: 1247-1248
Automated dermatological diagnosis: hype or reality?.
J Invest Dermatol. 2018; 138: 2277-2279
Performance of intensive care unit severity scoring systems across different ethnicities in the USA: a retrospective observational study.
Lancet Digit Health. 2021; 3: e241-e249
Dissecting racial bias in an algorithm used to manage the health of populations.
Science. 2019; 366: 447-453
Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations.
Nat Med. 2021; 27: 2176-2182
CheXclusion: fairness gaps in deep chest X-ray classifiers.
arXiv. 2020; ()
Radiology “forensics”: determination of age and sex from chest radiographs using deep learning.
Emerg Radiol. 2021; 28: 949-954
Artificial intelligence algorithm improves radiologist performance in skeletal age assessment: a prospective multicenter randomized controlled trial.
Radiology. 2021; 301: 692-699
Prediction of systemic biomarkers from retinal photographs: development and validation of deep-learning algorithms.
Lancet Digit Health. 2020; 2: e526-e536
Assessment of patient specific information in the wild on fundus photography and optical coherence tomography.
Sci Rep. 2021; 118621
Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning.
Nat Biomed Eng. 2018; 2: 158-164
Identifiability, exchangeability, and epidemiological confounding.
Int J Epidemiol. 1986; 15: 413-419
Confounding and collapsibility in causal inference.
SSO Schweiz Monatsschr Zahnheilkd. 1999; 14: 29-46
Equity in essence: a call for operationalising fairness in machine learning for healthcare.
BMJ Health Care Inform. 2021; 28e100289
Current clinical applications of artificial intelligence in radiology and their best supporting evidence.
J Am Coll Radiol. 2020; 17: 1371-1381
FDA cleared AI algorithms.
The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database.
NPJ Digit Med. 2020; 3: 118
The State of radiology AI: considerations for purchase decisions and current market offerings.
Radiol Artif Intell. 2020; 2e200004
Shades of difference: theoretical underpinnings of the medical controversy on black/white differences in the United States, 1830–1870.
Int J Health Serv. 1987; 17: 259-278
The biological concept of race and its application to public health and epidemiology.
J Health Polit Policy Law. 1986; 11: 97-116
MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports.
Sci Data. 2019; 6: 317
CheXpert: a large chest radiograph dataset with uncertainty labels and expert comparison.
AAAI. 2019; 33: 590-597
Measures of body composition in blacks and whites: a comparative review.
Am J Clin Nutr. 2000; 71: 1392-1402
Mammographic breast density and race.
AJR Am J Roentgenol. 2007; 188: 1147-1150
Heart disease and African Americans.
Disparities in cardiovascular disease risk in the United States.
Curr Cardiol Rev. 2015; 11: 238-245
Racial differences in bone density between young adult black and white subjects persist after adjustment for anthropometric, lifestyle, and biochemical differences.
J Clin Endocrinol Metab. 1997; 82: 429-434
Racial differences in bone strength.
Trans Am Clin Climatol Assoc. 2007; 118: 305-315
AI for radiographic COVID-19 detection selects shortcuts over signal.
Nat Mach Intell. 2021; 3: 610-619
Deep learning applied to chest X-rays: exploiting and preventing shortcuts.
PMLR. 2020; 126: 750-782
Recalibrating the use of race in medical research.
JAMA. 2021; 325: 623-624
Hidden in plain sight—reconsidering the use of race correction in clinical algorithms.
N Engl J Med. 2020; 383: 874-882
The algorithm audit: scoring the algorithms that score us.
Big Data Soc. 2021; ()
An algorithmic approach to reducing unexplained pain disparities in underserved populations.
Nat Med. 2021; 27: 136-140
Medical imaging algorithms exacerbate biases in underdiagnosis.
Research Square. 2021; ()
© 2022 The Author(s). Published by Elsevier Ltd.