Investigating Racial and Socioeconomic Characteristics in Pediatric Sepsis Using Electronic Health Data Free

This study was deemed exempt research by our institutional review board (institutional review board #20-018322). This was a retrospective study from a large, quaternary academic center, including sepsis episodes from January 20, 2011, to May 20, 2021, identified by an algorithm indicative of bacterial infection with concurrent organ dysfunction.

The sepsis algorithm was derived as part of an institutional program which links clinical and research data on >2 million patients. Infection was defined as a blood culture obtained (not necessarily positive) and administration of antibiotics for at least 4 days, similar to adult clinical definitions of sepsis.²⁵  This duration of antibiotics is meant to exclude patients for whom sepsis is suspected but “ruled out.” Organ dysfunction (Fig 1) was defined on the basis of criteria for pSOFA,⁹  with additional modifications taken from adult criteria and pediatric organ dysfunction score (PELOD 2 score).^11,12  Most importantly, this approach reflects the Sepsis-3 definition of sepsis, which reframes sepsis as infection with associated increase of ≥2 points on the SOFA organ dysfunction score.²⁶  Not only does this approach identify a broader cohort of patients, but it aligns with the efforts of pediatric medicine to mirror Sepsis-3.²⁷  During its validation period, this algorithm had 84% sensitivity, 65% specificity (positive prediction value of 43% and negative prediction value of 93%) when compared with manual chart review.⁸ 

All patients identified by algorithm above from January 20, 2011, to May 20, 2021, were included. No age limits were imposed, so as to fully represent the population seen at our center.

Patients with congenital heart disease admitted the cardiac center and NICU were excluded. For patients with >1 sepsis episode, we included only the most recent sepsis episode as their index episode. This approach ensured each patient was only represented once in the sample and maintained proximity to our primary outcome of mortality.

The primary exposures were race/ethnicity, social disorganization index (SDI), and insurance status. Race/ethnicity was categorized as non-Hispanic white (NH-white), non-Hispanic Black (NH-Black), Hispanic, and other (inclusive of patients identifying as Asian American, Native Hawaiian, Pacific Islander, American Indian, Alaskan Native or multiracial). Race and ethnicity were self-reported by patients and entered into the electronic medical record. Insurance status (at time of presentation) was categorized as private, public, and self-pay. The SDI was created on the basis of patient home zip code and incorporates the following socioeconomic parameters using the American Community Survey: overall unemployment, households receiving public assistance, low-income persons (<100% poverty level), low-income persons (100%–149% poverty level), high school dropouts, female-headed households, renter-occupied houses, and moved households within the last 3 years. These parameters were linked to zip codes on the basis of patient admission year, with admissions occurring in 2019 to 2021 linked to 2018 data because of data availability. To create the index, the parameters are summed and a z score is created on the basis of the sum. Patients were then stratified into high or low SDI, with low disorganization defined as quintiles 1 and 2 and high disorganization including quintiles 3 to 5. This index has been previously used to represent social determinants of health in previous literature.^28–31 

In addition, we examined relevant patient- and encounter-level covariates. Patient-level covariates included age, sex, time to hospital (on the basis of patient home zip code), and presence of complex chronic condition. Complex condition is coded using Feudtner et al’s complex chronic conditions categorization system, which refers to a comprehensive set of International Classification of Diseases codes that meet the following definition: “any medical condition that can be reasonably expected to last at least 12 months (unless death intervenes) and to involve either several different organ systems or 1 organ system severely enough to require specialty pediatric care and probably some period of hospitalization in a tertiary care center.”³²  Because there was no validated score of severity of illness across all patient settings (and because PELOD and pSOFA are intrinsically embedded in the sepsis criteria), we used number of organ dysfunctions on day 1 as a proxy. Organ dysfunction could be categorized as cardiovascular, respiratory, renal, or hematologic, referring to the surveillance definition used in the sepsis identification algorithm (Fig 1). At the encounter level, we looked at source of admission to hospital as presenting at the emergency department, ICU, inpatient floor, outside hospital, or perioperative care.

The primary outcome was all-cause mortality during index sepsis episode. Secondary outcomes included ICU admission during index sepsis episode, readmission within year after hospital discharge (for any indication), hospital-free days at day 30, ICU-free days at day 30, organ dysfunction-free days at day 30, and use of sepsis therapies (vasopressor, steroids, transfusions, or extracorporeal membrane oxygenation [ECMO]).

Standard descriptive statistics, n (%) and median (interquartile range [IQR]), were used to summarize patient demographics, mortality, and other secondary outcomes. Our study was powered to detect an odds ratio of at least 1.3 with respect to mortality (similar to previous studies).^23,24  Multivariable logistic regression was used to assess the association of race/ethnicity, insurance status, and SDI with in-hospital mortality, adjusted for age, sex, presence of at least 1 chronic complex condition, number of organ dysfunctions on day 1, source of admission, and time to hospital. Pearson’s χ² test and Fisher’s exact test were used to measure the association between our exposures and the following binary secondary outcomes: ICU admission, readmission within 1 year of hospital discharge, and use of sepsis therapies (vasopressor, steroids, transfusions, or ECMO). Kruskal-Wallis rank sum test and Wilcoxon rank sum test were used to examine association between exposures and hospital-free days at day 30, ICU-free days at day 30, and organ dysfunction-free days at day 30. Post-hoc tests for Fisher’s exact test and Dunn test for multiple comparisons were applied to those outcomes with P value ≤.05. Multiple comparison P values were adjusted with the Benjamini-Hochberg method. Statistical analysis was completed using SAS 9.4 (SAS Institute Inc., Cary, NC) and R (R core team 2019).

Among 7288 total sepsis episodes, there were 4532 unique patients represented. We subsequently described characteristics of each patient’s index episode (most recent episode), for which the mortality rate was 9.7% (Table 1). The median age was 7 (IQR, 2–14) years with 44% identified as NH-white, 28% NH-Black, 14% Hispanic or Latino, and 15% as other. Roughly half (48%) were privately insured, with the remainder publicly insured (<1% self-pay/charity care). A majority of patients (76%) had the presence of at least 1 complex chronic condition. Nearly half of index sepsis episodes were identified in the emergency department (45%), with the remainder occurring in the ICU (16%), inpatient floors (16%), outside hospital transfers (19%), and perioperative settings (3.7%). The average time to hospital (from patient zip code) was 41 minutes, with patients who died having longer travel time (51 minutes compared with 40 minutes, P value <.001).

Compared with NH-white, the unadjusted odds of mortality were significantly lower for NH-Black patients (0.69 [0.52–0.91]), but not significantly different for Hispanic patients (0.76 [0.54–1.06]) and other races (1.21 [0.92–1.59]) (Table 2). There was no difference in odds of mortality based on insurance. Patients residing in zip codes with higher social disorganization had lower unadjusted odds of mortality (0.78 [0.63–0.98]). However, once adjusted for age, sex, presence of at least 1 complex chronic condition, organ dysfunction on day 1, and source of admission, there was no significant association between exposures and mortality. An additional adjusted model which included time to hospital also showed no significant association with mortality.

Regarding admission rates, 68% required ICU admission during the sepsis episode and 50% were readmitted (for any reason) within 1 year (Table 3). NH-Black patients and those residing in zip codes with higher SDI were more likely to be admitted to the hospital (P value <.001 for both, Table 3). There was no significant association between our predictors and ICU admission. Hispanic patients and publicly insured patients were more likely to be readmitted within 1 year (Hispanic, 1.28 [1.06–1.55]; public, 1.19 [1.05–1.35]), whereas patients classified as other or self-pay were less likely to be readmitted (other, 0.79 [0.66–0.96]; self-pay, 0.43 [0.19–0.9]).

With respect to hospital-, ICU-, and organ dysfunction-free days (Table 4), NH-Black patients had significantly higher hospital-free and organ dysfunction-free days when compared with NH-white patients (NH-Black median hospital-free days, 20 [IQR, 7.7–25.2] compared with NH-white days, 19 [IQR, 0–24.8], P value = .011; NH-Black median organ dysfunction-free days, 27 [IQR, 23–29] compared with NH-white days, 27 [IQR, 20–29], P value = .035). Hispanic patients had lower organ dysfunction-free days when compared with NH-white patients (Hispanic median organ dysfunction-free days, 26 [IQR, 18–29] compared with NH-white days, 27 [IQR, 18–29], P value = .035).

When investigating the association of our predictors with additional sepsis therapies (Table 5), there was significantly less vasopressor use in NH-Black and Hispanic patients (compared with 32.9% NH-white; NH-Black, 27.5%, P value = .005; Hispanic, 28%, P value = .041), as well for those residing in zip codes with high social disorganization category (compared with low SDI 33.6%; high SDI 29.3%; P value = .003). There was also significantly less transfusion use in publicly insured patients compared with private (public, 26.8%, P value = .007). There were no significant differences in ECMO and steroid use among our predictors.

Overall, this is a novel approach to study disparities in a cohort of clinically defined pediatric sepsis, which leverages the power and information of the electronic health record. The mortality rate of 9.7% is comparable to previous studies using billing/administrative codes.^1,2,4  In contrast to previous studies,^23,33  we did not observe a difference in mortality for Black or Hispanic/Latino patients. In fact, these patients had lower odds of mortality, though these were not statistically significant after adjusting for age, sex, complex chronic condition, organ dysfunction on day 1, source of admission, and time to hospital. Rather than presume that disparities do not exist, we thought it was important to highlight the community context of our patients and practice location, as well as acknowledge that our findings may not be generalizable to other geographic areas. Since the racial composition of our city is different from that observed in this cohort (namely that the county is ?45% NH-Black and 34% NH-white,³⁴  compared with our cohort of 28% and 44%, respectively), we wondered if our findings were skewed by sicker patients in surrounding counties (which are predominately NH-white) that were referred to our quaternary care center. That said, we adjusted for complex chronic condition and number of organ dysfunctions on day 1 to account for potential differences in severity of illness. It is also possible that the lower unadjusted odds of mortality for NH-Black and Hispanic patients is because of proximity to the hospital. It is well known in sepsis that rapid fluid resuscitation and antibiotics are keys to survival; it may be that these patients were closer to the hospital (that additionally specializes in pediatrics), leading to quicker recognition and resuscitation. For this reason, an additional adjusted model including time to hospital was created (Table 2). This third model was nearly identical to the previous adjusted model, suggesting that time to hospital did not contribute significantly to mortality.

Additionally, we observed lower odds of mortality in patients residing in zip codes with high social disorganization, though this did not reach statistical significance when adjusted for age, sex, complex chronic condition, organ dysfunction on day 1, source of admission, and time to hospital. This is likely because of colinearity between race/ethnicity and SDI. Similar to the racial trends described above, we wondered if the patients with high social disorganization were representative of our urban environment and, again, the sicker patients were referred from the suburban areas surrounding the city. Along the same lines, these may be patients with higher medical complexity (>75% of this cohort had 1 or more complex chronic conditions) for which our center was their medical home.

When investigating admission, ICU admission, and readmission trends, it is important to note that almost all patients required admission, conditional on the algorithm definition of sepsis including 4 or more days of antibiotics. The remaining patients either died before admission, were transferred elsewhere, or left against medical advice. An important limitation of the surveillance definition (Fig 1) is that patients who had a blood culture and died before admission were included as a sepsis episode, though in many cases the true cause of death may be unknown (ie, cardiac arrest, sudden infant death syndrome, etc), but a blood culture was ordered as part of broad workup.

With regard to ICU admission, we did not observe any differences by race/ethnicity or SES. We did, however, observe that publicly insured and Hispanic or Latino patients were more likely to be readmitted within 1 year. This raised questions of whether postdischarge instructions, care, or follow-up is different for these groups. There are many previous studies which show increased readmissions for publicly or uninsured patients.^35–37  Conceivably, publicly insured patients may have different access to resources on the basis of coverage, or insurance status may just be a proxy for low SES and barriers to postdischarge care (inability to pay for medications, lack of transportation to follow-up visits, etc). It is also possible that medically complex or technology-dependent patients (who are more frequently readmitted) may be more likely to be publicly insured, though the data regarding insurance coverage of these patients is mixed.^38,39  With regard to the increased readmission of Hispanic patients, there may exist cultural or language barriers contributing to understanding of postdischarge care or anticipatory guidance, which may lead to readmission. However, ethnicity is poor proxy for English language proficiency and requires further investigation with a more specific variable representing English language proficiency. Previous studies have been mixed with regard to limited English proficiency as a risk factor for readmission.^40–42  We did not find increased readmission rate in Black patients, which has been previously documented.^36,43  It is likely that this trend was not observed because of the aforementioned racial composition of our single center.

When we examined secondary outcomes related to severity of illness, namely hospital- and organ dysfunction-free days (Table 4) and additional sepsis therapies (Table 5), we did observe that there was significantly less pressor use in Black and Hispanic, as well as in the high social disorganization group. This aligns with our previous observations that these patients may represent our local community-acquired sepsis, who are potentially less sick when compared with patients who are referred from other facilities.

There are several limitations to this study, including its single-center and retrospective nature. Nonetheless, our approach allows for investigation of clinical sepsis identified by electronic algorithm, decreasing the bias associated with billing/coding databases. This study is additionally limited by small sample size. Previous studies, among national databases, were able to show more subtle differences in mortality because of larger sample size, which may be missed in our smaller group. Despite this limitation, our study was powered to detect an odds ratio of at least 1.3 with respect to mortality (similar to previous studies)^23,24  and yields findings that can inform institutional and local practice. Lastly, we did not have a reliable indicator of severity of illness to show whether there are differences in how sick minority or low SES children are, which may be a more appropriate outcome than mortality. That said, we did incorporate number of organ dysfunctions on day 1 and ancillary sepsis therapies (vasopressor use, ECMO, etc), as well as hospital- and organ dysfunction-free days, which can shed light on the severity of sepsis.

Previously observed racial/ethnic and socioeconomic disparities in mortality were not confirmed when sepsis was identified using electronic clinical data; however, these results are not reflective of local or national trends because of the single-center nature of the study. There were significantly higher hospital readmissions in patients who were publicly insured or identify as Hispanic or Latino, which require further investigation. This study illustrates an innovative approach to investigating racial/ethnic and socioeconomic disparities in pediatric sepsis, by capitalizing on an algorithm which identifies clinically defined sepsis. This approach alleviates bias embedded in retrospective studies using billing/coding data and can be used to identify populations with disparate outcomes, which can be used to inform local and/or institutional health equity efforts.