Disparities in healthcare are widely recognized, especially regarding discrimination based on race and ethnicity (1, 2). Such disparities can unveil themselves as differences in quality of care, unequal medical device performance, or access to services reflecting structural inequities (3). These biases are not only harmful for patient care, but can also impact the development of machine learning-based clinical algorithms that train on electronic health records (EHR) (4).

Ensuring the development of fair AI models is crucial, and addressing missing information is a key initial step in achieving this objective, especially when such information is not missing at random (5, 6). Unfortunately, this variation in the level of monitoring is often not taken into consideration in the development of machine learning-based clinical algorithms. In a 2017 study that evaluated 107 electronic health record (EHR)-based risk prediction tools, 49 did not account for missing data (7). A common approach to imputation is the use of normal values based on the assumption that laboratory tests that are not ordered are presumed to be within normal range, a practice that likely introduces bias (8).

The probability of detecting an abnormal finding is contingent on the frequency of testing. Consequently, non-randomly missing data can lead to spurious correlations—non-causal relationships between features and outcome—that are learned and then incorporated into clinical algorithms (9). When the etiology of missing data stems from social determinants of care, these biases can become ingrained in subsequent AI models, perpetuating, and even scaling existing disparities (10, 11). This is even more important in a high-stake environment such as in patients with sepsis admitted to the intensive care unit (ICU).

Sepsis is a severe life-threatening systemic infection and effective management of this condition requires prompt diagnosis, aggressive treatment, and continuous monitoring. Despite current advances, one key challenge remains the timely delivery of care. Herein, serum lactate level is one of the two key diagnostic tools of septic shock according to the guidelines (12, 13). Disparities in sepsis outcomes are known to exist (14). However, the drivers of sepsis disparities are unknown and the question of whether disparities extend to serum lactate monitoring remains underexplored.

This paper seeks associations between race and ethnicity, sex, and language and the frequency of serum lactate determination during the management of sepsis in the ICU. By shedding light on this dimension of care, we aim to contribute to a more comprehensive understanding of the social patterning of the data generation process in healthcare.

This observational retrospective study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement (15). The health equity language, narrative, and concepts of this paper follows the American Medical Association's recommendations (16).

In this retrospective cohort study in patients with sepsis, we observed no discernible disparities between sexes and non-native English speakers in receiving a serum lactate measurement on day 1, although Black patients had a slightly increased likelihood. Furthermore, no apparent racial or language disparities were evident when examining the frequency of subsequent measurements, although a lower frequency was observed for women, those with private insurance, and those admitted electively. As Non-white patients were more likely to have Medicaid, there might still be disparities in care not captured in our data.

Although our study does not directly involve AI model development, its findings are highly relevant to the growing use of clinical data in artificial intelligence applications. Variability in measurement frequency, such as with lactate, can introduce biases into model training and deployment, particularly if the data reflects healthcare process differences rather than true physiological states. This is especially important given recent concerns about fairness and generalizability in AI models, which often underperform in underrepresented patient populations due to uneven data quality and representation (25, 26). Understanding and quantifying these real-world data characteristics is therefore, essential for building equitable and reliable AI systems in critical care. Health equity has become a priority in clinical research and among policymakers not only in the US but globally (27–29). In recent years, significant legislative changes around AI and health equity outcomes have been proposed and implemented. The European Parliamentary Research Service conducted a study on AI in healthcare in 2022 and recommended the implementation of specific coordination and support programs to address issues pertaining to AI and bias (30). In December 2023, the European Union approved the world's first legislation to regulate AI (31).

Beyond the obvious risks associated with feeding non-representative data to a model, variation in the clinical monitoring of patients presents a problem in the development of prediction, classification and optimization models using real-world data. The non-random sparsity of data from minoritized groups, even when represented in the dataset, has implications in the application of any statistical model. Machine learning-based decision support tools are an especially delicate area due to the sensitive nature of clinical decisions. Providers often intentionally refrain from measuring a variable especially in the ICU because of increasing recognition of the harm from over-testing (32). But the rationale behind such decisions is typically more complex, and confounded by both clinical and non-clinical (i.e., social determinants of care) features. In result, AI models learn wrong associations between clinical features and outcomes of interest. The problem becomes more pronounced in the advent of multi-modal modeling that requires black box deep learning representations (9). Models built on real-world data are thus subject to the human biases of the people who collected the primary data. For instance, a recent study found that large language models recommended low paying jobs more frequently to Mexicans, or implied that administrative work is solely a female job (33).

While our study did not assess mortality outcomes directly, partly due to the publication of similar work (34), the relationship between measurement frequency and patient outcomes remains an important area for future research. Causal inference methods, such as target trial emulation (35), could help determine whether more frequent monitoring leads to improved outcomes. In an effort to mitigate biases, some studies have suggested the use of causal inference frameworks for machine learning (33, 36, 37), which should help understand and avoid embedding biases into AI algorithms. Evaluating data inputs used in AI models for biases and disparities as done in our work is a prerequisite even before employing causal inference frameworks and should become standard practice as the understanding gained aids in building better, more equitable, and trustworthy AI models. This study provides a framework and approach for future work, as health care professionals, engineers, and developers have the moral accountability to ensure safe deployment of AI models (38, 39).

Lactate is frequently measured as part of bundled panels, such as point-of-care arterial blood gas analyses leading to synchronous measurement with other parameters. However, it can also be assessed independently in the central laboratory. In a previous study on blood glucose monitoring in the ICU, we compared point-of-care and central lab measurements and found no clinically relevant differences between the two methods (40), suggesting that the impact of measurement context may be limited in this setting. While our analysis was conducted using data from the MIMIC database, similar findings have been reported in other critical care datasets, including eICU and Duke. For instance, Matos et al. (41) observed patterns of variability in arterial blood gas measurements across these cohorts. This cross-dataset consistency suggests that our conclusions may be generalizable to other ICU populations and settings. The optimal frequency of monitoring of serum lactate measurement is unknown. Several recent studies and reviews have shown that serial lactate measurements and trends—such as peak levels, area under the curve, and clearance are associated with mortality in sepsis. For example, a 2023 nationwide Korean cohort study found that combining serial lactate values with SOFA improved mortality prediction compared to SOFA alone (42). A meta-analysis with roughly 4,400 patients suggests that a protocol focusing on lactate clearance leads to lower in-hospital mortality compared to ScvO2 normalization or usual care (43). A 2024 study in septic shock patients also linked initial, peak, and final 24-h lactate levels to 28-day mortality, although 24-h clearance was not predictive and optimal measurement frequency remained unclear (44). Similarly, a 2022 study using MIMIC-IV reported that lactate peak and AUC were associated with mortality, but lactate clearance was not more predictive than single values and performed worse than SOFA or NEWS scores (34). The current Surviving Sepsis Campaign guidelines support serial lactate monitoring but acknowledge that evidence for improving patient-centered outcomes remains limited and inconsistent (13). Many EHRs have already incorporated automated sepsis alerts to clinicians which rely on data such as the lactate to be present; disparities in collecting the data leads to disparities in usage of such alerts (45, 46). As such, the inputted data must be evaluated for bias. Other studies have shown that racially diverse Non-White ICU patients have nearly double the incidence of sepsis and higher rates of sepsis-related mortality compared to White patients (45, 47, 48). Furthermore, some studies in pediatric patients have reported higher mortality rates for those of lower socioeconomic status in the ICU (49). As such, all possible efforts need to be undertaken to close this disparity in patient care.

While our research provides valuable insights into the discourse on disparities and biases within critical care, it is essential to acknowledge the limitations of our study. Firstly, selection bias could be a potential concern, as our data only encompassed patients admitted to the ICU in an academic tertiary care center in the USA whose patients are predominantly White. However, data from MIMIC is generally very similar to data from eICU-CRD another publicly available database encompassing 208 ICUs in the US (50). In general, race-ethnicity is self-reported in MIMIC-IV or provided by relatives, however in instances where this was not possible, data is recorded by the providers themselves. Additionally, our study design precludes us from testing for unmeasured confounding variables, especially leading to confounding by indication. Future research endeavors should make concerted efforts to address these limitations, such as including Social Determinants of Health and fostering a more comprehensive understanding of the topic by employing causal inference frameworks as the next prerequisite step before validating AI models. Although our observed variability in lactate measurements was statistically significant, the absolute magnitude was small, mostly due to the large sample size, and may not be clinically meaningful. However, even small differences can carry relevance in the critical care setting, where clinical decisions often hinge on marginal changes. Recent studies have highlighted the challenges of bias and data completeness in electronic health records, particularly in underrepresented populations, which can impact both model performance and clinical interpretation (25, 26, 51). Our analysis focused on lactate due to its clinical relevance in sepsis and its frequent measurement in routine care, which allowed for robust assessment of intra-patient variability, but we agree that we cannot exclude incidental findings due to random variations. Also, similar disparities may affect other laboratory parameters (41), and our findings may not be generalizable to those. Future studies should explore a broader range of biomarkers to assess the extent of this issue across different clinical contexts. Moreover, future studies should extend their scope to cover other facets of care, including emergency departments, regular wards, or ambulatory care, to provide a more holistic perspective.

The implications of our study extend beyond the realm of lactate monitoring during sepsis management. In addition to the ongoing challenge of achieving healthcare equity within a system marked by systemic biases, clinicians and researchers must remain cognizant of these disparities before endeavoring to enhance patient care at their local institution or constructing any AI model. These biases not only have the potential to distort predictions, but may also endanger patient's safety when the predictions are employed for treatment or management decisions.