Development of a Clinical Prediction Model for Central Line-Associated Bloodstream Infection in Children Presenting to the Emergency Department

Abstract

Objective:

The majority of the children with a central line who present to the emergency department with fever or other signs of bacteremia do not have a central line-associated bloodstream infection (CLABSI). Our objective was to develop a clinical prediction model for CLABSI among this group of children in order to ultimately limit unnecessary hospital admissions and antibiotic use.

Methods:

We performed a nested case-control study of children with a central line who presented to the emergency department of an urban, tertiary care children’s hospital between January 2010 and March 2015 and were evaluated for CLABSI with a blood culture.

Results:

The final multivariable model developed to predict CLABSI consisted of 12 factors: age <= five years, black race, use of total parenteral nutrition, tunneled central venous catheter, double lumen catheter, absence of other bacterial infection, absence of viral upper respiratory infection symptoms, diarrhea, emergency department temperature >= 39.5 degrees Celsius, fever prior to presentation, neutropenia, and spring/summer season. The clinical prediction score had good discrimination for CLABSI with a c-statistic of 0.81 (CI: 0.77–0.85). A cut point less than 6 was associated with a sensitivity of 98.5% and a negative predictive value of 99.2% for CLABSI.

Conclusions:

We were able to identify risk factors and develop a clinical prediction model for CLABSI in children presenting to the emergency department. Once validated in future study, this clinical prediction model could be used to assess the need for hospitalization and/or antibiotics among this group of patients.

Introduction

Central lines are commonly used for the delivery of medications, nutritional support, and blood products for outpatient children with a variety of conditions. It is well known, however, that central lines pose certain risks.ⁱ Central line-associated bloodstream infections (CLABSIs) are the most common complication of central lines and often contribute to a number of serious adverse clinical events.ⁱⁱ As a result, CLABSIs have a significant impact on healthcare spending with costs of up to $40,000 per infection.ⁱⁱⁱ

Because of this risk, non-hospitalized children with central lines often undergo an evaluation for CLABSI when they present to the emergency department with symptoms or signs of possible bacteremia, such as fever or lethargy.^iv,v Most hospitals admit and empirically treat these patients with broad-spectrum antibiotics for up to 48 hours while awaiting blood culture results.^vi,vii Because the consequences of delaying treatment for CLABSI are often severe,^viii it is important that infections be identified early and with high sensitivity. Fever, however, is a common occurrence in the pediatric population and is not specific for CLABSI. Therefore, the majority of children with central lines presenting with fever are not bacteremic.^ix These children are not only admitted to the hospital unnecessarily but are also exposed to broad-spectrum antibiotics, which are associated with their own risks.^x

Previous studies have identified risk factors for CLABSI in hospitalized, critically-ill patients, and other specific patient populations, including fever, use of total parenteral nutrition, and double lumen catheters,^{xi,xii,xiii,xiv,xv} but no studies have identified a generalized set of risk criteria to categorize outpatients with central lines at risk for CLABSI. Identifying such a set of criteria could allow clinicians to select children in whom hospitalization and antibiotics could be avoided, and who could instead be discharged from the emergency department with close monitoring at home. Similar to the work to identify children at low risk for bacterial meningitis performed by Nigrovic and colleagues,^xvi we sought to develop a tool to guide decision making in the population of pediatric outpatients with a central line presenting with symptoms of infection in order to better classify risk for CLABSI and to ultimately limit unnecessary hospital admissions and antibiotic use.

Our specific objective in this study was to develop a clinical prediction model for CLABSI among children with a central line presenting to the emergency department with symptoms of infection.

Materials and Methods

We conducted a retrospective, nested case-control study at the Children’s Hospital of Philadelphia (CHOP) between January 1, 2010 and March 5, 2015. The CHOP Institutional Review Board approved this study. Data for this study were obtained via extraction from electronic health records and chart review.

Study Cohort

We defined our study cohort as all pediatric patients (ages 0 to 18 years) with a central line who presented to the CHOP emergency department with symptoms consistent with possible bacteremia (as determined by clinical assessment of the emergency department physicians). Operationally we defined this as all patients who had a blood culture sent from their central line within a window of 12 hours prior to emergency department arrival (to include those patients who had a blood culture drawn by an outpatient clinic before being referred to the emergency department) to 24 hours after emergency department arrival. Patients were excluded from the cohort if they had been admitted to the hospital within the seven days prior to presentation or if they had received parenteral antibiotics before their initial blood culture was drawn. For patients with multiple emergency department evaluations for possible bacteremia during the study time period, only the first presentation was included for formal analysis.

Cases

Our primary outcome was laboratory-confirmed bloodstream infection in a patient with a central line, as defined by the Centers for Disease Control and Prevention National Healthcare Safety Network (NHSN).^xvii We identified cases with a positive blood culture using the CHOP Ambulatory CLABSI Database, which has been maintained since January 2010. This database was started as a CHOP quality improvement initiative and prospectively captures all outpatients diagnosed with a CLABSI after presenting to CHOP. Cases of secondary CLABSI (as defined by NHSN) are excluded from the database.

Controls

We randomly selected control patients with negative blood cultures from the same cohort as the case patients. We matched them to cases in a 2:1 ratio based on year of presentation and presence of an oncologic diagnosis.

Potential Risk Factors

The risk factors collected on each patient fell under the broad categories of patient demographics, central line characteristics, baseline health status, and signs and symptoms on presentation (Table 1). We selected variables to evaluate based on their potential association with CLABSI as determined by literature review and hypothesized associations. Based on results from prior publications and the distribution of the data, we dichotomized several variables: ‘number of days central line is in place’ (<= 14 days vs. > 14 days);^xviii ‘emergency department temperature’ (< 39.5 degrees Celsius vs. >= 39.5 degrees Celsius);^xix and ‘age’ (<= 5 years vs. > 5 years). We defined ‘season of presentation’ as either Spring/Summer (April 1 through August 31) or Fall/Winter (September 1 through March 31) based on the typical seasonality of respiratory viral activity in the Philadelphia area. We divided white blood cell count into less than 4,000, 4,000 to 18,000, and greater than 18,000 cells per high-powered field.

Table 1.

Univariable Analysis for Association with CLABSI

Variable	Cases, No. (%)	Controls, No. (%)	Odds Ratio	P value	95% CI
Baseline Patient Characteristics
Age <= 5 years	108 (51)	152 (36)	2.1	<0.001	1.4–2.9
Black Race	49 (23)	63 (15)	1.6	0.03	1.1–2.4
Male	117 (55)	228 (54)	1.1	0.70	0.8–1.5
Patient Health Status
Hospital Admission in Past 3 Months	170 (81)	272 (64)	2.5	<0.001	1.7–3.8
Oncologic Diagnosis	108 (52)	223 (52)		matched
Systemic Antibiotics	38 (18)	69 (16)	1.2	0.50	0.8–1.8
Systemic Steroids	37 (18)	72 (17)	1.0	0.90	0.6–1.6
Total Parenteral Nutrition	50 (24)	24 (6)	6.1	<0.001	3.1–11.9
Central Line Details
Central Line Days > 14	189 (91)	339 (86)	1.6	0.14	0.9–2.9
Central Line Typea
PICC	15 (7)	70 (17)	1.0	reference
Port	78 (37)	277 (67)	1.2	0.58	0.6–2.7
Tunneled CVC	103 (49)	53 (13)	9.8	<0.001	4.4–21.9
Double Lumen	87 (41)	60 (15)	4.0	<0.001	2.6–6.1
Prior CLABSI in Past 3 Months	23 (11)	14 (3)	3.9	<0.001	1.9–8.0
Recent Break in Line or Dressingb	15 (7)	15(4)	1.9	0.08	0.9–3.9
Sign of Infection at Insertion Site	24 (11)	21 (5)	2.0	0.03	1.1–3.8
Signs and Symptoms on Presentation
Absence of Viral URIc Symptoms	151 (72)	240 (56)	2.0	<0.001	1.4–2.9
Absence of Other Bacterial Infection	202 (96)	363 (85)	3.6	0.001	1.7–7.4
Diarrhea	43 (20)	56 (13)	1.8	0.008	1.2–2.9
ED Temperature >= 39.5 Celsiusd	27 (13)	19 (4)	2.7	0.001	1.5–5.1
Emesis	64 (30)	98 (23)	1.5	0.03	1.0–2.3
Fever Prior to Presentatione	178 (84)	371 (79)	1.6	0.05	1.0–2.5
White Blood Cell Count < 4	74 (37)	146 (35)	1.1	0.67	0.7–1.6
White Blood Cell Count > 18	15 (7)	26 (6)	1.4	0.40	0.7–2.8
Neutropenicf	50 (25)	94 (23)	1.2	0.44	0.8–1.9
Spring/Summer Seasong	91 (43)	145 (34)	1.5	0.03	1.0–2.1

^aPICC = peripherally inserted central catheter, CVC = central venous catheter;

^bWithin the prior 30 days;

^cUpper respiratory infection;

^dFirst recorded emergency department temperature;

^eMeasured or tactile fever as reported by caregiver;

^f< 500 neutrophils per high powered field;

^gApril 1 – August 31

The patient’s guardian self-selected the patient’s race from the hospital provided options of ‘American Indian or Alaska Native,’ ‘Asian,’ ‘Black or African-American,’ ‘Native Hawaiian or Other Pacific Islander,’ ‘White,’ ‘More Than One Race,’ ‘Other,’ and ‘Refused.’ We defined neutropenia as an absolute neutrophil count <= 500 cells per high-powered field. We defined ‘recent break in line or dressing’ as any caregiver report of breach in integrity of the central line itself or the sterile dressing covering the line entrance site within the 30 days prior to presentation.^xx,xxi We defined ‘fever prior to presentation’ as a measured temperature greater than 38 degrees Celsius or tactile “fever” as reported by caregiver in the 24 hours prior to presentation according to chart documentation. ‘Viral upper respiratory infection symptoms,’ ‘emesis,’ and ‘diarrhea’ were determined by caregiver report of the 24 hours prior to presentation as well as any signs or symptoms noted by the emergency department physician upon initial assessment. Initial emergency physician physical exam was used to determine ‘signs of infection at line insertion site.’ ‘Other bacterial infection’ was determined by caregiver report of a current bacterial infection diagnosed by a physician (such as acute otitis media diagnosed by primary care physician) or by a new diagnosis made by the emergency physician prior to discharge from the emergency room (such as urinary tract infection or bacterial pneumonia).

Several variables of interest were not consistently documented in the electronic medical record or not collected on the majority of the patients (e.g. use of ethanol or antibiotic locks, material of central line, vascular location of line, and c-reactive protein level) and therefore were excluded from our analysis.

Statistical Analysis

Overall rate of positivity for CLABSI was calculated for all evaluations as well as for the cohort if only a patient’s first presentation (incident presentation) was included. The rate of positivity for CLABI per year was tested for trend using an extension of the Wilcoxon rank-sum test. We assessed potential predictor variables for univariate associations with CLABSI using conditional logistic regression. We then created a multivariable prediction model using backwards stepwise conditional logistic regression to eliminate variables while minimizing the Akaike Information Criterion in order to obtain a more parsimonious model, as recommended by the TRIPOD guidelines.^xxii To create a clinical prediction score, we divided each beta coefficient in the final model by the smallest beta coefficient and then rounded to the nearest integer.²² The clinical prediction score for each patient was the sum of these integers. To assess model performance, we evaluated the c-statistic of the model and the sensitivity, specificity, and negative predictive value of the clinical prediction scores with the aim of optimizing sensitivity and negative predictive value. All analyses were performed using Stata 13.1.

Results

There were a total of 3,226 eligible evaluations for suspected CLABSI during the study period, of which 403 (12.5%) met our case definition of CLABSI (Table 2). Stratified by presence or absence of oncologic diagnosis, the number of evaluations that were confirmed CLABSI was 237 (9.4%) for patients with an oncologic diagnosis and 166 (23.2%) for non-oncologic patients. When assessing only incident presentations during the study period, there were 1,798 incident presentations and 227 incident cases of CLABSI leading to a rate of positivity for CLABSI of 12.6% and oncologic and non-oncologic cases of 123 (8.4%) and 104 (30.6%) respectively. There was a significant trend in decreasing rate of CLABSI per year when all presentations are included, for both the entire cohort (beginning with a rate of 20% in 2011 and ending with a rate of 6.7% in 2015) (P = .03) and for the separate oncologic (P = .04) and non-oncologic groups (P = .04).

Table 2.

Annual Incidence Rates of Positivity for CLABSI - All Presentations

Year	Total Presentations (No.)	CLABSI Negative (No.)	CLABSI Positive (No.)	Rate of Positivity (%)	Rate of Positivity Oncologic (%)	Rate of Positivity Non-Oncologic (%)
2010	509	407	102	20.0%	12.0%	47.0%
2011	737	642	95	12.9%	9.8%	22.7%
2012	642	561	81	12.6%	11.0%	18.8%
2013	592	531	61	10.3%	8.6%	16.3%
2014	627	571	56	8.9%	6.9%	16.7%
2015a	119	111	8	6.7%	5.6%	10.0%
Overall	3226	2823	403	12.5%	9.4%	23.2%

^aJan 1, 2015 – March 5, 2015

Of the 227 incident cases of CLABSI during the study period, we excluded 27 due to lack of complete data, leaving 200 complete cases for analysis. These cases were matched to 399 controls (for one case it was only possible to find one matching control). The study group was 46% female and 54% male. The group was made up of 52% oncologic patients and 48% non-oncologic patients. Several variables were significantly associated with CLABSI in univariable analysis (Table 1).

Our final multivariable model consisted of 12 variables: age less than or equal to 5 years, black race, total parenteral nutrition use, absence of other bacterial infection, absence of viral upper respiratory illness symptoms, diarrhea, first recorded emergency department temperature greater than or equal to 39.5 degrees Celsius, fever reported prior to presentation, neutropenia, spring/summer season, tunneled central venous catheter, and double lumen catheter (Table 3). Prediction scores were assigned to each of the variables. Total possible prediction score ranged from 0 to 26.

Table 3.

Final Multivariable Analysis for Association with CLABSI

Variable	OR	P Value	95% CI	Risk Score
Baseline Patient Characteristics
Age <= 5 yrs	1.7	0.04	1.0–2.8	1
Black Race	2.1	0.03	1.1–4.2	2
Patient Health Status
Total Parenteral Nutrition	4.8	<0.001	2.0–11.5	3
Central Line Details
Tunneled CVCa	5.1	<0.001	2.5–10.2	4
Double Lumen	2.4	0.01	1.2–4.5	2
Signs and Symptoms on Presentation
Absence of Other Bacterial Infection	2.4	0.06	1.0–5.9	2
Absence of Viral URI Symptomsb	2.7	0.001	1.5–4.8	2
Diarrhea	2.0	0.04	1.0–4.0	2
ED Temperature >= 39.5 Celsiusc	2.2	0.08	0.9–5.3	2
Fever Prior to Presentationd	2.0	0.04	1.0–3.9	2
Neutropenice	3.3	<0.001	1.7–6.4	3
Spring/Summer Seasonf	1.6	0.08	0.9–2.7	1
Maximum Possible Prediction Score				26

^aCVC = central venous catheter;

^bUpper respiratory infection

^cFirst recorded emergency department temperature;

^dMeasured or tactile fever as reported by caregiver;

^e< 500 neutrophils per high powered field;

^fApril 1 – August 31

Black race was an unexpected risk factor identified in our final multivariable model. To further investigate this we looked at a conditional logistic regression of black race for CLABSI with payer status (Medicaid v. private insurance) added as a possible modifying variable. In this simple regression with payer status added, race was no longer significantly associated with CLABSI.

Performance of Clinical Prediction Model

The clinical prediction score had good discrimination for CLABSI in this derivation cohort, with a c-statistic of 0.81 (95% confidence interval: 0.77–0.85) (Figure 1). The median clinical prediction score for cases was 11 and for controls was 7. Sensitivity, specificity, and negative predictive value were determined for each score (Table 4). A cut-point of a score of 6 was associated with a sensitivity of 98.5% and a negative predictive value of 99.2% for CLABSI.

Fig 1.

Receiver-operating characteristic (ROC) curve and corresponding c-statistic for the CLABSI risk score.

Table 4.

Performance of Clinical Prediction Score

Cut-point	Sensitivity	Specificity	Negative Predictive Value
>= 2	100.0%	0.0%	-
>= 3	99.5%	2.8%	97.7%
>= 4	99.5%	3.8%	98.1%
>= 5	99.0%	16.0%	99.0%
>= 6	98.5%	28.6%	99.2%
>= 7	93.5%	43.4%	97.9%
>= 8	85.5%	58.4%	96.6%
>= 9	74.0%	69.9%	94.9%
>= 10	66.0%	77.4%	94.1%
>= 11	56.0%	84.7%	93.1%
>= 12	48.0%	90.5%	92.4%
>= 13	38.0%	94.5%	91.4%
>= 14	26.5%	96.2%	90.2%
>= 15	19.0%	98.3%	89.5%
>= 16	10.5%	99.0%	88.6%
>= 17	5.5%	99.3%	88.0%
>= 18	2.5%	99.8%	87.8%
>= 19	2.0%	99.8%	87.6%
>= 20	1.0%	99.8%	87.5%
>= 21	0.5%	100.0%	87.5%

We performed a sensitivity analysis to assess the model performance separately in the oncologic and non-oncologic patients and found that the model performed somewhat better in oncologic patients than in non-oncologic patients (c-statistic of 0.85 v. 0.76). We also performed the analysis limiting the study group to only those patients who had a fever prior to presentation and found good performance of the prediction model in this group with a c-statistic of 0.81.

Discussion

Through our study we identified 12 risk factors for CLABSI in outpatient children, and from these risk factors we constructed a clinical prediction model that had good discrimination for CLABSI. In our analysis of the baseline rates of positively for CLABSI, we found that the rates of CLABSI for non-oncologic patients were much higher than for oncologic patients. We speculate that this may be due to the fact that oncology patients are much more likely to have a fully implanted central line (port) instead of another type of central line. Ports have been shown to be a protective factor for CLABSI.^xxiii In our study group, 77% of oncology patients had ports whereas only 34% of non-oncology patients had ports. When including all presentations, the rate of CLABSI decreased each year between 2010 and 2015, possibly representing the positive effects from continued improvements in outpatient CLABSI prevention such as maintenance care bundles and antibiotic locks.^{xxiv,xxv,xxvi,xxvii}

Our final model included several risk factors that have previously been associated with CLABSI in various subsets of our population: fever,^xxviii double lumen catheter,^xxix total parenteral nutrition,^xxx tunneled central venous catheter,^xxxi and younger age.⁹ We also identified several novel risk factors: lack of viral upper respiratory symptoms, spring/summer season, absence of other bacterial infection, diarrhea, and black race.

Black race was significantly associated with CLABSI on both univariable and multivariable analysis, with an odds ratio of 2.1 (CI: 1.1–4.2) in the final multivariable model. It is unclear what the ultimate source of this association is. There have been recent publications looking at the increased rate of healthcare associated infections in African Americans,^xxxii,xxxiii although whether the patients themselves are different or the care they receive is different remains unknown. Because we lacked any direct measures of socio-economic status in our dataset, it is possible that race in our model functions as a proxy for socio-economic status and that lower socio-economic status could be associated with increased risk for CLABSI, perhaps due to more instability in the home, decreased education of the caregivers caring for the line, or a general lack of resources available for the care of the patient. Our finding that black race is no longer associated with CLABSI when payer status is added into a basic conditional logistic regression supports this hypothesis.

There were several limitations to our study. First, the number of cases was small and prevented us from being able to perform an internal validation using a temporally distinct sample. We instead plan to perform validation in a separate cohort of patients. Second, due to missing data we were unable to include certain risk factors in our analysis, such as use of ethanol or antibiotic locks. Third, because we performed matching based on oncologic status, we were unable to include this as a possible risk factor for CLABSI. We chose to do this because we believe oncologic patients may ultimately behave differently in terms of CLABSI risk, and we wanted to make sure this factor did not drive the findings of our model to the exclusion of other risk factors. And in fact, the clinical prediction model appears to perform differently in oncologic and non-oncologic patients. We will approach validation of the clinical prediction model separately in non-oncologic and oncologic patients.

Long-term central lines in non-hospitalized patients can improve patient survival as well as patient and family quality of life. To maximize quality of life, however, adverse events associated with these central lines must be minimized. Significant progress has been made in preventing ambulatory CLABSIs, but to maximize the quality of life of patients living at home with central lines it is important not only to prevent CLABSIs, but also to decrease the number of times these patients are unnecessarily admitted to the hospital and exposed to broad spectrum antibiotics. It is our hope that, once validated, this clinical prediction model could assist physicians in determining the appropriate and safe disposition for outpatient pediatric patients at risk for CLABSI. For example, patients with a clinical prediction score of less than 6 could be deemed at very low risk for CLABSI and be discharged home to await blood culture results, while patients with a clinical prediction score of 6 to 8 could be deemed as intermediate risk and receive a dose of ceftriaxone and be sent home, and patients with higher scores could be deemed as high risk and be admitted as is the current standard of care. Ultimately, this could enable clinicians to spare some of these patients the life disruption of hospital admission, limit exposure to broad-spectrum antibiotics, and decrease healthcare spending.

Conclusion

We were able to identify risk factors and develop a clinical prediction model for CLABSI in children presenting to the emergency department. The model performed well among oncologic and non-oncologic patients. Future study should validate this model. Once validated in a separate population, this clinical prediction model could be used to assess for risk of CLABSI and potentially decrease unnecessary hospital admissions among this group of patients.