Self-reported Race/Ethnicity and Intraoperative Occult Hypoxemia: A Retrospective Cohort Study

After receiving Institutional Review Board approval from the Icahn School of Medicine at Mount Sinai (New York, New York), data were extracted from our departmental data warehouse for all adult patients (older than 18 yr) who received an anesthetic that included at least one arterial blood gas (ABG) sample between January 2008 and December 2019. Our departmental data warehouse is a unified data source that contains data from both our historical anesthesia information management system (CompuRecord; Philips Medical Systems, USA) as well as our electronic health record (Epic; Epic Systems, USA). Matching of historical intraoperative anesthesia information management system data to additional demographics from the electronic health record, and import of these additional data into our departmental data warehouse, is done via an automated process and was not specifically undertaken for this project. Informed consent was waived by the institutional review board because of the retrospective nature of the study.

For each Sao₂ concentration calculated by ABG (GEMStat Premier 3000; Instrumentation Laboratory, USA), the corresponding Spo₂ was found by calculating the mean Spo₂ during a 5-min interval starting 10 min before the ABG time (interval ended at 5 min before the ABG time). This interval was chosen to represent the prevailing Spo₂ values at the time the laboratory specimen was likely drawn, as well as to account for specimen transport time before logging/processing of the specimen using point-of-care ABG analyzers. ABG data were excluded if there were no comparative Spo₂ values available during this interval, or if Sao₂ or Spo₂ values were greater than 100% and deemed to be device error. Spo₂ values were captured every 15 s for all cases. The calculated Sao₂ value and calculated Spo₂ value were then linked to one another as a “paired Sao₂–Spo₂ reading” for the purpose of data analyses.

Pulse oximeter device bias was calculated as the difference between the averaged Spo₂ and Sao₂ values. Occult hypoxemia was defined as Sao₂ less than 88% despite a Spo₂ of greater than 92%, based on previously published parameters.¹⁶  Our analysis included the full range of Spo₂ values between 92 to 100%.

Patient- and case-level data extracted included basic patient demographics (age, sex, body mass index, self-identified race, self-identified ethnicity); smoking status at time of surgery (current, previous, or never smoker); year of procedure as a categorical variable; ASA Physical Status; history of relevant comorbidities (hypertension, chronic pulmonary disease, congestive heart failure, coronary artery disease, valvular heart disease, diabetes, peripheral vascular disease, renal failure); anesthesia type (general, monitored anesthesia care, regional anesthesia); use of volatile anesthetic agent (yes/no); use of vasoactive infusion (yes/no [for use of a phenylephrine, norepinephrine, epinephrine, or vasopressin infusion at any time during the portion of the anesthetic]); mean arterial pressure (MAP); hematocrit value as reported at the time of the associated ABG; and set ventilator parameters (fraction of inspired oxygen [Fio₂], tidal volume, positive end-expiratory pressure [PEEP]). Measured ventilatory parameters were not recorded or saved to the departmental data warehouse. For each case, we calculated the mean Fio₂, tidal volume, and PEEP for the same interval that was used to calculate mean Spo₂. The year of procedure was included to account for changes in pulse oximeters used at our institution (primarily Nellcor [Medtronic, USA] before 2011 and Masimo [Masimo, USA] thereafter). Comorbidities were calculated from International Classification of Diseases, Ninth Revision and Tenth Revision codes using the icd package¹⁷  in R (R Foundation for Statistical Computing, Vienna, Austria). This is an open-source, validated package that assigns comorbidities using the methods of Elixhauser et al.¹⁸  and Quan et al.¹⁹ 

Self-reported race and ethnicity data collected at time of admission were used for our analyses. Options for self-reported race and ethnicity varied during the time of the study period but were cross-referenced for agreement. Race and ethnicity data were classified into a synthetic race/ethnicity variable according to a previously described method.²⁰  Specifically, we modified the U.S. Census Bureau/Office of Management and Budget (Washington, D.C.) standards for classification of race/ethnicity to assign each patient to one of the following groups: White (European, North African, Middle Eastern, non-Hispanic), Black (African American, sub-Saharan African, non-Hispanic), Hispanic/Latinx, Asian, or Other race/ethnicity.^21,22  Race/ethnicity was synthesized into a single variable in order to include Hispanic/Latinx patients who may be difficult to categorize using standard U.S. Census guidelines, as a growing percentage of these patients identify race as “Other”²³ ; thus, all patients who self-identified as Hispanic ethnicity were categorized into the Hispanic race/ethnicity group. Patients self-identifying as Native American/Alaskan Native, Indian/South Asian, and Pacific Islander were grouped into the Other race/ethnicity group due to an extremely low number of these patients in our dataset, which caused instability of the multivariable model during preliminary analysis. The full process used to assign self-reported race/ethnicity responses to the single synthetic race/ethnicity variable is outlined in Supplemental Digital Content 1 (https://links.lww.com/ALN/C797).

No statistical power calculation was conducted before the study since the sample size was based on available data in our data warehouse. The statistical and data analysis plans were defined before accessing the data and were finalized after the data were accessed, with additional statistical analysis completed after peer review. Patient and case characteristics were compared across self-reported race/ethnicity using one-way ANOVA for continuous variables that met parametric assumptions. Continuous variables that did not meet parametric assumptions, including nonnormal distributions as assessed by histograms, were analyzed using Kruskal–Wallis tests. The chi-square test was utilized for categorical variables. Two-tailed testing was used for all statistical tests. The reference group for race/ethnicity was White; the reference group for sex was female. For ASA Physical Status, ASA Physical Status I and II were grouped together because of the small number of ASA Physical Status I patients (less than 1.5%); this combined ASA Physical Status I and II group was used as the reference.

Generalized estimating equations modeling was utilized to determine if any race/ethnicity was an independent predictor of occult hypoxemia relative to White race/ethnicity within the study Spo₂ range of 92 to 100%. A Poisson regression model with logarithmic link was used because of the low prevalence of occult hypoxemia. Multiple paired Sao₂–Spo₂ readings for each patient were accounted for by clustering within patients using independent working correlation structure; however, sandwich estimators were utilized for standard errors. The model was first run without controlling for confounders in order to assess for significant differences in unadjusted prevalence of occult hypoxemia across race/ethnicity groups. A multivariable model was then created to control for all of the collected demographic, comorbidity, and operative variables. A continuous Spo₂ variable was included as a fixed effect, and an interaction term between Spo₂ and race/ethnicity was also assessed for inclusion. Continuous variables (e.g., Fio₂, tidal volume) were entered directly into the model as covariates, without transformation. The year of procedure was assessed as a categorical variable; 2011 was used as the reference level (since this was the approximate year that the pulse oximeter device type changed). All covariates included were assessed as confounding variables rather than effect modifiers. An additional model focusing on 92 to 96% Spo₂ was created as a supplemental analysis, as this was the range focused on in the analysis of occult hypoxemia from Sjoding et al.¹⁶ 

As an exploratory analysis, the optimal threshold for predicting occult hypoxemia based on Spo₂ was assessed using the F₁ score. The F₁ score is the harmonic mean of the precision (positive predictive value) and recall (sensitivity), and is a measure of a test’s accuracy.²⁴  It can be used to determine the threshold at which a test has the best performance, particularly in cases where the outcome (in this case, occult hypoxemia) is rare. F₁ scores were calculated using univariate analysis for the range of Spo₂ between 92 to 100%, stratified by race/ethnicity, and then plotted. This allowed for visual identification of the best Spo₂ cutoff to identify occult hypoxemia.

The overall design of the analysis and threshold for statistical significance (P < 0.05) were established a priori. After reviewing our findings, several post hoc sensitivity analyses were performed to further elucidate the relationship between race/ethnicity, occult hypoxemia, and comorbid risk factors.

All statistical analysis was performed in R version 3.5.0 (R Foundation for Statistical Computing).

Readings from 47,067 unique patients were collected, totaling 157,482 Sao₂–Spo₂ pairs. Before data analysis, 17 paired Sao₂–Spo₂ readings were excluded because of missing Sao₂ data, and an additional 28 paired Sao₂–Spo₂ readings were excluded for Sao₂ values greater than 100, deemed to be device error. There were 11,232 paired Sao₂–Spo₂ readings for patients whose race was recorded as “Unknown,” who were grouped into the Other race/ethnicity group. We then isolated paired Sao₂–Spo₂ readings to the Spo₂ 92 to 100% range, which left a total of 151,070 paired Sao₂–Spo₂ readings for 46,253 unique patients. Demographics for the cohort for data analysis are listed in table 1. The most represented race/ethnicity groups by number of unique patients were White (22,089 of 46,253 [47.8%]), Other (10,071 of 46,253 [21.8%]), Hispanic (6,304 of 46,253 [13.6%]), and Black (5,177/46,253 [11.2%]). The White cohort was older (61 vs. 52 to 54 yr) than the non-White groups on average, and had a greater proportion of male sex (12,530 of 22,089 [56.7%]) compared to the Black (2,381 of 5,177 [46%]) and Hispanic (3,220 of 6,304 [51.1%]) cohorts. The proportions of male sex in the Asian (1,526 of 2,612 [58.4%]) and Other (5,569 of 10,071 [55.3%]) cohorts were similar to the White cohort. The Black, Hispanic, and Other cohorts had greater proportions of ASA Physical Status IV and V, diabetes, and renal failure compared with the White and Asian cohorts. Rates of congestive heart failure, hypertension, and chronic pulmonary disease were similar between the White, Black, and Hispanic cohorts, greater in the Other cohort, and lesser in the Asian cohort. Rates of general anesthesia were high across all groups, with few study patients having received monitored anesthesia care or regional anesthesia. The median Spo₂ was 100% (interquartile range, 99 to 100) and median calculated Sao₂ was 100% (interquartile range, 100 to 100) for the overall cohort as well as for each race/ethnicity group.

The prevalence of occult hypoxemia in the Spo₂ study range for each self-reported race/ethnicity group is listed in table 2 and illustrated in figure 1. Overall, occult hypoxemia was present in 1.3% (2,016 of 151,070) of paired Sao₂–Spo₂ readings. The prevalence of occult hypoxemia in Black patients was 2.1% (339 of 16,011 paired Sao₂–Spo₂ readings) versus 1.1% (791 of 70,722 paired Sao₂–Spo₂ readings) in White patients (P < 0.001). Hispanic patients also had a significantly greater prevalence of 1.8% (383 of 21,223 paired Sao₂–Spo₂ readings; P < 0.001). Asian and Other race/ethnicity had a similar prevalence to White race.

The overall Spo₂ device bias was 0.0 ± 6.8% (mean ± SD). White race/ethnicity was the only group that demonstrated a negative Spo₂ device bias (−0.2 ± 6.3%). Positive Spo₂ device bias was observed for Black (0.6 ± 9.1%), Hispanic (0.5 ± 7.9%), Asian (0.2 ± 6.5%), and Other (0.1 ± 5.9%) patients.

Factors associated with occult hypoxemia in the multivariable analysis for the Spo₂ study range are shown in table 3. Black (odds ratio, 1.44 [95% CI, 1.11 to 1.87]; P = 0.006) and Hispanic (odds ratio, 1.31 [95% CI, 1.03 to 1.68]; P = 0.031) race/ethnicity were associated with significantly greater odds of occult hypoxemia relative to White race/ethnicity. Higher Spo₂ was associated with decreased risk of occult hypoxemia (odds ratio, 0.71 [95% CI, 0.69 to 0.73]; P < 0.001). The overall interaction term between race/ethnicity and Spo₂ was not significant (P = 0.948) and was thus not included in the final model. ASA Physical Status V was an independent risk factor for occult hypoxemia (odds ratio, 1.83 [95% CI, 1.16 to 2.88]; P = 0.009). Higher PEEP (odds ratio, 1.10 [95% CI, 1.06 to 1.13]; P < 0.001) and higher hematocrit (odds ratio, 1.05 [95% CI, 1.03 to 1.06]; P < 0.001) were also associated with a significantly greater risk of occult hypoxemia. In contrast, higher tidal volume (odds ratio, 0.80, [95% CI, 0.75 to 0.84]; P < 0.001) and higher MAP (odds ratio, 0.89 [95% CI, 0.84 to 0.95]; P < 0.001) were associated with significantly lower risk of occult hypoxemia. Older age was associated with lower odds of occult hypoxemia (odds ratio, 0.94 [95% CI, 0.89 to 0.99]; P = 0.012). A more recent year of procedure (2015 to 2019) was also associated with decreased odds of occult hypoxemia compared with the reference year (2011; table 3).

Several post hoc sensitivity analyses were performed based on the results of the multivariable analysis. Specifically, because of the strong association between ASA Physical Status V and occult hypoxemia, as well as the greater proportion of ASA Physical Status V patients in the Black, Hispanic, and Other groups, multivariable analyses were repeated while excluding ASA Physical Status V patients. The association between Black (odds ratio, 1.40 [95% CI, 1.08 to 1.82]; P = 0.011) and Hispanic (odds ratio, 1.29 [95% CI 1.01 to 1.64]; P = 0.039) race/ethnicity and occult hypoxemia was still significant with comparable odds ratios. Full results of this sensitivity analysis are shown in Supplemental Digital Content 2 (https://links.lww.com/ALN/C797). We also performed a sensitivity analysis limited to the 92 to 96% Spo₂ range, as per Sjoding et al.¹⁶  Results were consistent with our primary analysis. Black race/ethnicity was significantly associated with occult hypoxemia (odds ratio, 1.80 [95% CI, 1.10 to 2.94]; P = 0.020) versus White race/ethnicity, with a greater odds ratio than in the full 92 to 100% Spo₂ model. The association of occult hypoxemia in Hispanic patients no longer reached significance using the Spo₂ range of 92 to 96% (odds ratio, 1.46 [95% CI, 0.95 to 2.25]; P = 0.082). The full results of this analysis can be found in Supplemental Digital Content 3 (https://links.lww.com/ALN/C797).

The F₁ score was calculated for White, Black, and Hispanic patients. F₁ scores for Asian and Other race/ethnicity were not calculated given the lack of significant difference in occult hypoxemia prevalence relative to the White reference group. Supplemental Digital Content 4 (https://links.lww.com/ALN/C797) includes a figure with graphic representation of the calculated F₁ scores. The peak F₁ score for White patients was at Spo₂ 94% (F₁ score, 0.08), meaning that Spo₂ values less than 94% were most associated with occult hypoxemia. The F₁ score for Black race/ethnicity peaked at 96% (F₁ score, 0.11). The peak F₁ score was at 95% for Hispanic race/ethnicity (F₁ score, 0.13).

In this large, retrospective, single-center analysis, we found that self-reported Black race/ethnicity (339 of 16,011 paired Sao₂–Spo₂ readings [2.1%]; odds ratio, 1.44 [95% CI, 1.11 to 1.87]) and Hispanic race/ethnicity (383 of 21,223 paired Sao₂–Spo₂ readings [1.8%]; odds ratio, 1.31 [95% CI, 1.03 to 1.68]) were significantly associated with occult hypoxemia in the 92 to 100% Spo₂ range compared with White race/ethnicity (791 of 70,222 paired Sao₂–Spo₂ readings [1.1%]). The interaction between Spo₂ and race/ethnicity was not significant, supporting our finding that race/ethnicity is associated with occult hypoxemia independent of Spo₂. These findings are consistent with recent descriptions of occult hypoxemia in the critical care population,¹⁶  but contrary to previous studies suggesting racial discrepancies exist at Spo₂ values less than 80% and are minimal at more commonly encountered Spo₂ values.^14,15  The sensitivity analysis of the 92 to 96% Spo₂ range ([Black race/ethnicity] odds ratio, 1.80 [95% CI, 1.10 to 2.94]) is consistent with the findings of Sjoding et al. The lack of statistical significance of Hispanic race/ethnicity versus White race/ethnicity in the 92 to 96% analysis is likely attributed to underpowered analyses.

Patients with ASA Physical Status V also demonstrated a significant association with occult hypoxemia (odds ratio, 1.83 [95% CI, 1.16 to 2.88]) across all race/ethnicity groups, potentially because of derangements in physiology or therapeutic interventions associated with patients of ASA Physical Status V. Although a larger percentage of ASA Physical Status V patients were in the Black and Hispanic groups, sensitivity analyses demonstrated this relationship remained significant with the exclusion of ASA Physical Status V patients in both Black (odds ratio, 1.40 [95% CI, 1.08 to 1.82]) and Hispanic (odds ratio, 1.29 [95% CI, 1.01 to 1.64]) groups.

Surprisingly, neither the use of vasoactive infusions nor the diagnoses of peripheral vascular disease or diabetes were associated with occult hypoxemia despite previous literature indicating an association.^25–27  Sjoding et al. excluded diabetic patients from their analysis, so there is no recent and direct comparison to our study available.¹⁶  The associations between age, ventilatory parameters, MAP, and hematocrit with occult hypoxemia were all statistically significant; however, associations were hard to interpret clinically since the magnitude of the effect size was generally very small and often in a contradictory direction (e.g., higher PEEP appeared to be associated with more occult hypoxemia; table 3). Further evaluation of the impact of these covariates is required across all populations and is beyond the scope of this work. There is no clear explanation as to why year was significantly associated with occult hypoxemia, as it does not correspond to the change in pulse oximeter manufacturers in 2011.

Despite our primary findings, the overall Spo₂ device bias in White patients compared to non-White patients was not as striking on univariate analysis, with a difference of less than 1.0% across all groups. The pulse oximeter slightly underestimated Sao₂ in White patients and slightly overestimated Sao₂ in non-White patients. It is likely that Spo₂ device bias does not occur in all non-White patients, which may explain the discrepancy between the increased prevalence of occult hypoxemia in the Black and Hispanic self-reported populations, but more similar prevalence of overall Spo₂ device bias. This discrepancy may be attributed to the wide range of skin pigmentations in patients who self-identify as Black and Hispanic, including multiracial individuals who self-identify as Black or Hispanic because of historical racial integrity laws.^28,29 

As shown in our primary analysis, lower Spo₂ (evaluated as a continuous variable) is significantly associated with a increased risk of occult hypoxemia. The univariate F₁ scores presented in Supplemental Digital Content 4 (https://links.lww.com/ALN/C797) were calculated to explore if there are optimal Spo₂ thresholds for predicting occult hypoxemia. Based on our findings, it may be prudent to have a greater index of suspicion for occult hypoxemia if Spo₂ is less than 96% for Black patients and 95% for Hispanic patients (compared with 94% for White patients). The magnitude of the F₁ scores was not high regardless of race/ethnicity (range, 0.08 to 0.16), indicating the ability of Spo₂ in the 92 to 100% range to reflect hypoxemia is low. This is not surprising since the overall prevalence of hypoxemia in that range is very low and occult hypoxemia is an unexpected event. Thus, the difference in peak F₁ scores between each race/ethnicity should not be overinterpreted.

Limitations of this study are important to consider, given the potential clinical impact of our findings. The retrospective nature of the study contributes to several limitations including lack of specific self-reported race/ethnicity (hence, the Other race/ethnicity group) for a large portion of our sample and the use of two different manufacturers’ pulse oximeters during the study period, as well as other unknown patient-specific confounding factors such as jaundice, presence of nail polish, skin thickness, probe placement, and probe location. The large sample size of this study likely limits the impact of most individual patient factors, but the overall small number of patients found to have occult hypoxemia should be considered a limitation. Additionally, the contribution of each race/ethnicity to the overall number of paired Sao₂–Spo₂ readings compared with the number of patients is slightly different, which potentially indicates selection bias due to repeated values for the same individual. However, this risk for selection bias is limited because of the very small differences in these populations (table 1). To help account for differences in the number of paired Sao₂–Spo₂ readings per patient, we utilized a generalized estimating equations model with clustering within patients. Additionally, we must recognize our use of “race/ethnicity” is a surrogate measurement for skin pigmentation. There is likely wide variation in skin pigmentation among the races/ethnicities represented in our data that cannot be addressed given the retrospective nature of the study.

The variety of pulse oximeters used at our institution during the study period (primarily Nellcor before 2011; Masimo thereafter) was controlled for using the year of procedure as a covariate in the multivariable analysis—although this is an imperfect proxy. It is possible that pulse oximeters from other manufacturers have better or worse performance with differing skin pigmentations, thus limiting the generalizability of our study. Additionally, unknown concentrations of carboxyhemoglobin may exist in this population, but we believe this risk is mitigated by demonstrating a lack of association between smoking status and occult hypoxemia in the multivariable model, as well as the fact that patients suffering from thermal injury are not cared for at our institution. The risk of carboxyhemoglobinemia is also likely mitigated given our institution’s adherence to Anesthesia Patient Safety Foundation (Rochester, Minnesota) guidelines on the use of modern carbon dioxide absorbents.³⁰  Calculated—rather than measured—Sao₂ presents a potential limitation and may limit comparisons to the Sjoding et al. study in which cooximetry was used to directly measure Sao₂.

Despite this, previous studies evaluating the ABG analyzer used during the duration of the study demonstrate accuracy of the Sao₂ calculation.^31,32  Body temperature could potentially affect pulse oximetry data and has not been included as a covariate, which must be recognized as a limitation of this study. The clinical indication for ABG sampling is also unknown given the retrospective nature of this study. Finally, anesthesia type may affect our findings: because of the low sample size of the regional anesthesia and monitored anesthesia care cases compared with general anesthesia cases, our model was unable to adjust for anesthesia type as a covariate.

In conclusion, we have found evidence that Black and Hispanic race/ethnicity are significantly associated with occult hypoxemia in the 92 to 100% Spo₂ range in anesthetized patients. Although the lack of objective skin pigmentation information and the retrospective nature of this study may limit the strength of our findings, we believe it is important to recognize the limitations of monitoring devices and maintain vigilance to avoid unrecognized hypoxemia.

Support was provided solely from institutional and/or departmental sources.

The authors declare no competing interests.