Introduction

Racial differences in the delivery of clinical care disproportionately harm racial minorities. For example, multiple studies have found lower rates of colon cancer screenings and increased all-cause mortality among Black patients [1]. AI models could amplify biases present in the data, further hindering health equity. In particular, AI models may rely on spurious correlations or confounded relationships to predict outcomes of interest [2,3]. For example, AI models trained on data that reflect racial health disparities (e.g., differences in healthcare spending) under-recommend Black patients for high-risk care management programs [4]. Without accounting for spurious correlations between race and the clinical outcome of interest, AI models may encode and amplify racial inequities in the delivery of clinical care.

Racial differences in laboratory testing rates are a potentially widespread source of bias in AI models for healthcare settings. Many AI models are developed to predict clinical outcomes in which the pragmatic definition depends on laboratory test results (e.g., sepsis [5,6]), such that untested patients are assumed to have a normal result. This assumption is often operationalized by assigning untested patients a negative label during model training [7–12]. However, under this assumption, if laboratory testing rates differ across race, AI models may exhibit disparate performance across racial subgroups [13]. Such models could inappropriately underestimate risk for patients in racial groups less likely to receive laboratory tests, potentially amplifying inequities in clinical care. While previous work [14,15] has examined racial differences in specific laboratory tests in patient subpopulations, to the best of our knowledge, a large-scale study of racial differences in laboratory testing has not been performed.

We performed a retrospective cohort study of laboratory testing differences between Black and White patients visiting the emergency department (ED) at two large U.S. teaching hospitals. We analyzed the ED since it features a wide range of populations and clinical conditions with ample diagnostic data. Furthermore, patients are not pre-selected for care and generally do not choose their providers. ED patient cohorts are also commonly used in training AI models for predicting clinical outcomes [7,9,16–18]. We quantified testing rates within Black and White patients across seven common laboratory tests with and without matching on age, biological sex, chief complaint, and ED triage score.

Methods

Results

After applying exclusion criteria (Fig 1), we analyzed 336,824 and 541,310 ED visits at BIDMC and U-M, respectively (Table 1). At BIDMC and U-M, Black patients accounted for 92,437 (27.4%) and 102,313 (18.9%) ED visits, respectively. In the unmatched cohort, White patients were significantly older than Black patients (BIDMC: 55 vs. 46 years, difference: 9 years, p < .001; U-M: 52 vs. 43 years, difference: 9 years, p < .001) and were less often female (BIDMC: 52.0% vs. 62.0%, 95% CI: -10.5% to -9.3, p < .001, U-M: 53.1% vs. 57.0%, 95% CI: -4.5% to -3.5%, p < .001). The ED triage score distribution also differed significantly across White and Black patients (p < .001). On average, White patients were assessed as more ill than Black patients (i.e., lower ED triage score; BIDMC: 2.6 vs. 2.8, difference: -0.2; U-M: 2.7 vs. 2.8, difference: -0.1). At BIDMC and U-M, 157,520 (46.8%) and 196,918 (36.4%) patients were admitted to the hospital following the ED visit, respectively. White patients were admitted at significantly higher rates than Black patients (BIDMC: 51.0% vs. 36.0%, difference 15.0%, 95% CI: 14.3% to 15.5%, p < .001; U-M: 37.9% vs. 29.8%, difference 8.1%, 95% CI: 7.6% to 8.6%, p < .001). We also report the 25 most common chief complaints by race at each institution (S1 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Cohort diagrams for BIDMC and U-M.

Cohort diagram of patient groups selected for study at BIDMC (top) and U-M (bottom). Note that at BIDMC, there were no ED visits in BIDMC associated with individuals of unknown sex. Numbers may not sum to 100.0% due to rounding.

https://doi.org/10.1371/journal.pgph.0003555.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Cohort demographics for BIDMC (left) and U-M (right) among all ED visits for the unmatched cohort.

https://doi.org/10.1371/journal.pgph.0003555.t001

The unmatched analysis also found that White patients were significantly more likely than Black patients to receive all laboratory tests in our analysis except d-dimer (Fig 2, top). For example, White patients were significantly more likely than Black patients to receive a CBC (BIDMC: 69.6% vs. 60.5%, difference 9.1%, 95% CI: 8.5% to 9.8%, p < .001). Fig 2 (top) shows similar findings for metabolic panels, blood cultures, ABG, troponin tests, and BNP (i.e., all other tests analyzed other than d-dimer). Trends were similar at U-M. See S2 Table for the testing rates shown in Fig 2 (top).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Differences in testing rates between White and Black patients, with and without matching on age, sex, chief complaint, and ED triage score.

Forest plots show difference in testing rate across White vs. Black patients (White %—Black %), with no matching adjustment (top), and matched on age, sex, chief complaint, and ED triage score (bottom). Differences at BIDMC are shown in red; differences at U-M are shown in blue. Actual values shown in S2 Table (results with no matching adjustment) and S4 Table (matched analysis).

https://doi.org/10.1371/journal.pgph.0003555.g002

Exact 1:1 matching on age, sex, ED triage score, and chief complaint yielded 94,320 and 141,510 ED visits at BIDMC and U-M, respectively. At BIDMC, this cohort included 51.0% and 19.3% of ED visits by Black and White patients, respectively. At U-M, this cohort included 69.2% and 16.1% of ED visits by Black and White patients, respectively. Matched cohort demographics can be found in S3 Table. Matching narrowed differences in testing rates (Fig 2). White patients remained significantly more likely than Black patients to receive a CBC (BIDMC: 63.0% vs. 61.3%, difference 1.7%, 95% CI: 1.1% to 2.4%, p < .001), metabolic panel (BIDMC: 63.2% vs. 61.7%, difference 1.5%, 95% CI: 0.9% to 2.1%, p < .001), or blood culture (BIDMC: 10.3% vs. 9.5%, difference 0.9%, 95% CI: 0.5% to 1.2%, p < .001), while White patients were significantly less likely than Black patients to receive a troponin test (BIDMC: 17.2% vs. 19.3%, difference -2.1%, 95% CI: -2.6% to -1.6%, p < .001). Differences in laboratory testing were not statistically significant at both institutions for BNP, d-dimer, or ABG. Trends were similar at U-M (Fig 2, bottom). See S4 Table for the testing rates shown in Fig 2 (bottom). White patients remained significantly more likely than Black patients to be admitted to the hospital after an ED visit at both institutions (BIDMC: 37.7% vs. 35.0%, difference 2.7%, 95% CI: 2.0% to 3.3%, p < .001; Fig 3, left). Trends were similar at U-M (Fig 3, right).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Differences in admission rates between White and Black patients matched on age, sex, chief complaint, and ED triage score.

Bar plots show difference in rate of admission to the hospital among ED visits in study cohort grouped by site. Differences at BIDMC are shown in red; differences at U-M are shown in blue. “**” denotes statistical significance at 5% level. Actual values are shown in S3 Table.

https://doi.org/10.1371/journal.pgph.0003555.g003

As a sensitivity analysis, we repeated the matched analysis without ED triage scores. Results were similar. White patients were significantly more likely than Black patients to receive a CBC, metabolic panel, and blood culture, and significantly less likely than Black patients to receive a troponin test (S1 Fig and S5 Table). White patients also remained more likely to be admitted to the hospital (S2 Fig). The analysis disaggregated by biological sex found similar trends: within male and female patients, White patients were significantly more likely than Black patients to receive a CBC or metabolic panel, and significantly less likely than Black patients to receive a troponin test (S3 Fig and S6 Table). Disaggregating by admission status narrowed some testing differences in visits resulting in discharge (S4 Fig and S7 Table) and visits resulting in admission (S4 Fig and S8 Table). However, given the significant differences in admission rate across race, narrower testing differences do not necessarily signal less systematic bias.

Discussion

Before presenting our findings, as an exercise in reflexivity, we as authors acknowledge how our positionality influenced this project. A subset of the authors serve as clinicians in the ED and the ICU at one of the hospitals in our analyses. Another subset of our authors are machine learning researchers. Several of the authors have prior work on health disparities research. In addition, one author had clinical training at a hospital that cared for a predominantly Black patient population. The authors identify as a White North American woman, White North American men, Asian Indian American man, Asian Taiwanese American man, and a female first-generation immigrant to the United States from Asia. None of the authors have lived experiences with anti-Black racism. It was our prior research in addition to the education received around implicit bias and racial disparities in access to clinical care that inspired, in part, the initial hypothesis.

We acknowledge that, as an observational study, our work has some extractive elements, since the authors do not interact with the communities from which data are sourced. Furthermore, our work is unlikely to immediately and directly remediate inequities in the delivery of clinical care, despite our alignment with equitable goals. However, using the power of our positionality as computer scientists and clinicians at a research institution, we aim to amplify cautionary messages about AI with respect to marginalized populations.

We now summarize the findings. Across six of seven laboratory tests, in the unmatched cohort, Black patients had significantly lower testing rates than White patients. After matching to account for differences in clinical characteristics, Black patients remained significantly less likely to receive a CBC or metabolic panel, while White patients were significantly less likely to receive a troponin or BNP test. Hospital admission partially explained racial differences in laboratory testing rates, with Black patients significantly less likely to be admitted to the hospital. Given similar findings across two institutions, these trends may hold more broadly.

Although this study is unable to examine the reasons for laboratory test orders, testing differences could impact AI models trained to predict outcomes defined using laboratory test results. For example, Black patients were found to be significantly less likely to receive blood cultures. They may subsequently be less likely to be diagnosed with serious infections. AI models trained to predict this outcome may underestimate risk in Black patients since models are generally trained to optimize a performance metric (e.g., accuracy) without explicitly accounting for confounding (i.e., between race and the outcome of interest). Specifically, such models could ultimately “learn” that Black patients are less likely to experience serious infections without “accounting” for lower admission rates and lower testing probability.

While quantifying differences in AI model performance across race under biased laboratory testing is not possible without “ground truth” diagnosis labels, studies in similar problem settings have consistently demonstrated the potential for harm. Past work has documented disparate performance in AI models with respect to race in healthcare applications [4,26]. Furthermore, a simulation study of disparities in testing across population subgroups suggests ample potential for performance gaps to emerge under biased laboratory testing patterns [13], such as those found in this study. Ultimately, acting on predictions of AI models trained on data exhibiting racial differences in laboratory testing could reinforce undertesting in marginalized patient populations. Furthermore, since laboratory tests are often used as a source of signal for a variety of downstream diagnoses, even these seemingly small differences could have widespread impacts on the equitable delivery of clinical care.

In general, machine learning researchers do not adjust for confounding prior to training ML models. Potential solutions for addressing selection bias are well-studied but may have limited practical utility for correcting biases in laboratory testing. For example, in the econometric and epidemiology literature, methods such as reweighting or matching are standard in observational studies to adjust for selection bias/survey sampling bias [27,28]. These approaches are most useful when there is no unobserved confounding, which is difficult to guarantee in practice. One may also consider removing race as an input to the model, but such approaches are insufficient if other variables are correlated with race (i.e., proxy discrimination [29,30]). In certain situations, it may be appropriate to train AI models on a sub-cohort where the spurious correlation is weaker (e.g., admitted or tested patients). However, this technique may discard most available training data, leading to lower model performance. Furthermore, the resulting model would not apply beyond the sub-cohort used for training (e.g., predicting serious infections in patients without a blood culture, using data from patients who received a blood culture) [17,31].

Thus, we require novel approaches to mitigate the potential impacts of racial testing differences on AI models. We highlight causally-motivated methods as a promising direction, since they could be used to model mechanisms of bias in laboratory testing and facilitate “counterfactual fairness” analyses (e.g., would a patient with these characteristics be tested regardless of their race) [32]. One may also consider extending techniques from semi-supervised/weakly-supervised learning (i.e., treating untested patients as unlabeled instead of negative), many of which assume that label missingness occurs randomly, motivating future methods that account for potentially biased laboratory testing patterns [33]. In parallel, it is imperative that we examine how racial testing differences emerge, to inform policies that minimize biases in laboratory testing, reducing the prevalence of spurious correlations in clinical data. Ultimately, due to the potential for AI to amplify existing harms, one should carefully consider how to adjust for racial differences in laboratory testing in AI models. A “color-blind” approach that ignores correlations between race and testing rates is unlikely to mitigate harm [34].

Limitations

Our study has several limitations. First, race is not an objective/unambiguous construct, which is reflected in our data. Some racial categories in BIDMC could overlap with patients who identify as either White or Black (e.g., “Brazilian,” “Hispanic/Latino”) or mixed-race individuals. We excluded such patients from our study. Our definition of “White” and “Black” captures the majority (79.6%) of all ED visits in BIDMC, mitigating the impact of such ambiguity. There may also be heterogeneity in testing differences within White and Black patients, which this study is unable to capture [35]. Second, to simplify the analysis, we restricted the study to White and Black patients. We recognize that there may be testing differences across other racial and ethnic groups. Past work highlights the disproportionate burden of racial disparities in clinical care on Hispanic/Latino, American Indian/Alaska Native, and Native Hawaiian/Pacific Islander populations in the U.S [36].

Third, matching on additional variables (e.g., comorbidities, additional markers of illness severity, patient history) could further explain correlations between race and testing rates. Such variables may also be subject to bias. Furthermore, exact-matching on additional variables could induce statistical bias if many units are dropped, while relaxations to exact matching may yield incomparable matched cohorts [37]. Fourth, our cohorts are drawn from U.S. academic teaching hospitals, which may not accurately reflect standard care across the U.S. Fifth, this study was unable to determine the appropriateness of testing rates. Although some of our findings align with previous work on the disproportionate impact of undertreatment/underdiagnosis on Black patients in various clinical contexts [1,15,38], the differences we observed could represent either under- or over-testing. Further study of racial differences in testing is warranted. Lastly, due to differences in data format, the laboratory test extraction pipelines necessarily differed across institutions. This could result in over/undercounting of tests. These limitations have minor impacts since our main outcome measure is testing differences rather than testing rates.

Conclusion

Black patients were observed to be significantly less likely than White patients to receive multiple laboratory tests across two large academic hospitals. Racial differences in testing narrowed or reversed after matching to account for differences in clinical characteristics. Hospital admission also partially explained these gaps. However, many racial testing differences remained significant. The findings highlight a need to better understand how racial differences in ED laboratory testing emerge, since the patterns observed in this study could represent either over- or under-testing. Furthermore, racial testing differences may amplify bias in AI models, since such models may spuriously associate race with laboratory test outcomes. Standard methods for addressing confounding may be insufficient to mitigate spurious correlations.

Supporting information