Predicting Serious Bacterial Infections Among Hypothermic Infants in the Emergency Department Free

We conducted a retrospective multicenter case-control study of infants presenting to 4 free-standing pediatric EDs between January 1, 2015 and December 31, 2019. Institutional review board approval was obtained from all participating sites. Data were uploaded to the Research Electronic Data Capture (REDCap) database (Vanderbilt University, Nashville, TN). We adhered to the guidelines described in the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis.¹⁰ 

We included infants ≤90 days of age presenting to the ED with any rectal temperature <36.5°C measured during their ED encounter and who had a blood culture drawn. We used this temperature cutoff based on the World Health Organization’s classification of hypothermia¹¹  as prior literature has been unable to identify a reliable cutoff for hypothermia to predict SBI.¹²  We retained only the first ED encounter for infants having multiple eligible encounters during the study period. We only included infants with a blood culture to limit our analysis to infants with clinical concern for SBI by the treating physician.

Our primary outcome was SBI. Bacteremia and meningitis were defined as growth of a pathogenic organism on blood or cerebral spinal fluid culture, respectively.¹³  We followed published criteria to classify suspected pathogens and contaminants, with ambiguous cases classified based on investigator consensus.^6,10  UTI was defined as growth from a catheterized urine culture of ≥50 000 Colony Forming Units/mL or 10 000 to 49 999 Colony Forming Units/mL with a positive urinalysis (defined as either ≥5 white blood cells [WBC] per high-power field, positive leukocyte esterase, or positive nitrite).⁶  Our secondary outcome was invasive bacterial illness (IBI), defined as bacteremia and/or bacterial meningitis.

Cases were defined as infants with a culture-confirmed SBI. Controls were defined as infants with hypothermia who had a blood culture drawn without SBI. We matched each case to 2 controls based on site and date of presentation.

We determined covariables a priori based on literature describing risk factors for infection in febrile and hypothermic infants.^5,6,14–18  Historical and physical examination components were obtained from the medical record via chart abstraction. Historical factors included chief complaint, age, sex, prematurity (defined as gestational age at birth <37 weeks), maternal historical factors (group B Streptococcus status, history of herpes simplex virus, intrapartum fever, and antibiotics received), history of NICU admission, and historical illness factors including seizure, abnormal behavior (irritability, poor feeding, or lethargy), hypothermia or pyrexia at home, respiratory distress, apnea, and cyanosis. Rash, seizure, and history of admission to the NICU were considered negative if not documented. Maternal historical data were obtained from any section of the medical record.

Physical examination covariables included weight, triage temperature, ED minimum and maximum temperature, ED minimum heart rate, ED minimum systolic blood pressure, ED minimum respiratory rate, abnormal examination, respiratory distress, apnea, and rash. Abnormal examination was defined using a consensus of terms for general ill-appearance descriptors (“sick,” “toxic,” “inconsolable,” “irritable,” and “meningismus”), compromised perfusion (“shock,” “decreased pulses,” “mottled,” “cyanotic,” “pale,” and “ashen”), mental status changes (“somnolence,” “lethargy,” “seizure,” and “altered”), and sunken or full fontanelle within the ED note.^14,19  Laboratory covariables included WBC count, absolute neutrophil count (ANC), platelet count, C-reactive protein, procalcitonin, abnormal urinalysis, glucose, total bilirubin, aspartate aminotransferase, alanine transaminase, respiratory virus positivity, and albumin.

We calculated descriptive statistics. Variables with a missing data proportion of ≥15% were excluded from multivariable analysis. For the remaining variables, we performed multiple imputation by chained equations.²⁰  We explored nonlinear associations between the probability of SBI and these quantitative parameters via restricted cubic splines. Following inspection of splined variables, age was converted to a dichotomous variable (≤21 and >21 days). WBC (≥12 000 cells per mm³), ANC (≥4500 cells per mm³), and platelets (≥328 000 cells per mm³) were dichotomized based on prior findings derived from this sample.²¹  ED vital signs of respiratory rate (<27 breaths per minute), heart rate (<130 per min), and blood pressure (<70 mm Hg) were dichotomized based on previously reported centile curves.^22,23  Minimum temperature was dichotomized to ≤36°C and >36°C.

We described the univariable association of each predictor with SBI using logistic generalized estimating equations models clustered by the matched case-control groups. We fit a logistic regression model including all predictors (the “full” model), penalizing the likelihood of mitigating the large number of predictors and optimizing the penalty by systematic search. To produce a simpler model, we performed variable selection using the Akaike Information Criterion and backward bootstrap resampling over 1000 iterations (the “reduced” model). Because of the large number of candidate variables, we derived a smaller model using 6 variables selected based on perceived clinical applicability (age, ED minimum temperature, fever, abnormal examination, elevated neutrophil count, and abnormal urinalysis). We expressed our results from each model as the odds ratio with 95% confidence interval (CI). We calculated the concordance index (c-index), a quality of fit measure for models that produce risk scores where a c-index of 1.0 equates to a perfect predictive model. We derived optimism-corrected c-indexes by bootstrap resampling of the model over 1000 iterations. We identified optimal thresholds for the predicted probability of SBI by analysis of the receiver-operator characteristic curve and calculated diagnostics metrics (sensitivity, specificity, and likelihood ratios) at the receiver-operator characteristic-identified optimal threshold.

We conducted 3 sensitivity analyses. First, we fit the full model with IBI as the outcome to identify candidate risk factors for IBI. Second, we fit the full model to a subset of well-appearing infants, defined as those who did not meet criteria for abnormal examination based on ED physical examination. Third, we evaluated risk factors in a subset of infants without fever (either historical or documented in the ED). Analyses were conducted using the MatchIt,²⁴ mice,²⁵ rms²⁶  packages in R, v4.2.2 (R Foundation for Statistical Computing, Vienna, Austria).

We identified 3516 infants with a rectal temperature of <36.5°C. Of these, 140 were repeat encounters for the same patient and were removed. Among 934 (27%) infants with a blood culture drawn, 57 had an SBI (1.7% of the total sample, 6.1% of infants with a blood culture). SBIs included 41 (72%) UTIs, 8 (14%) isolated bacteremia, 4 (7%) UTI with concomitant bacteremia, and 4 (7%) with meningitis (Table 1). The most common isolated organisms were Escherichia coli, Enterococcus faecalis, and group B Streptococcus.

The 57 infants with SBI were matched to 114 SBI-negative controls (Table 2). Prediction plots for quantitative predictors revealed a thresholding effect at 21 days and minimal nonlinearities with other predictors (Supplemental Fig 2). In univariable analyses, the following were positively associated with presence of SBI: age >21 days, fever at home, ED maximum temperature ≥ 38.0°C, elevated WBC, ANC, and platelets, and abnormal urinalysis. Prematurity, hypothermia at home, hypotension, and respiratory distress were negatively associated with SBI.

The full multivariable model had a raw c-index of 0.95 (95% CI: 0.92–0.97) and an optimism-corrected c-index of 0.91 (95% CI: 0.88–0.94; Table 3). Predicted probabilities of SBI from the full model exhibited good calibration with observed probabilities, with an average bootstrap-derived intercept of −0.01 and slope of 0.99 (Fig 1). The optimal cut point for the predicted probability was 0.32, which produced a sensitivity and specificity of 0.88 (95% CI: 0.76–0.95) and 0.89 (95% CI: 0.81–0.94), respectively.

Bootstrap backward selection selected only abnormal urinalysis for the reduced model, which was present in 97% of the fitted models. No other predictor was selected in >1% of models. The reduced model exhibited an optimism-corrected c-index of 0.82 (95% CI: 0.75–0.89). The odds ratio associated with abnormal urinalysis was 34.0 (95% CI: 13.5–85.4). The full and reduced models classified 7 (12%) and 16 (28%) infants as false negatives, respectively (Table 4). The smaller model using 6 variables chosen by consensus found that each factor was suggestive of SBI aside from abnormal examination and had an optimism corrected c-index of 0.91 (Supplemental Table 6).

Sixteen infants in our sample had IBIs (1.7% of included infants). In univariable analysis, respiratory distress was negatively associated with IBI, whereas elevated WBC was positively associated with this outcome (Supplemental Table 7). Although a full multivariable model for IBI exhibited a raw c-index of 0.85, no single variable was significantly associated with IBI in this adjusted model, and the optimism corrected c-index was 0.70. In our second sensitivity analysis, 84 (74%) controls and 45 (79%) cases were classified as well-appearing. Factors associated with SBI in well-appearing infants were similar to those from the full cohort analysis, as were c-indexes of fit (0.97 raw and 0.92 optimism-corrected, Supplemental Table 8). Our third sensitivity analysis limited to infants without fever included 22 cases (including all 4 infants with meningitis) and 94 controls. Within this sample, presence of abnormal urinalysis, leukocytosis, and elevated ANC were positively associated with SBI in univariable analysis (Table 5). Our multivariable model demonstrated that abnormal urinalysis was positively associated with SBI, whereas respiratory distress was negatively associated with this outcome.

We identified predictors for SBI in hypothermic infants presenting to the ED from a case control sample. The overall model performed well at predicting SBI, though with a high rate of false negatives, suggesting that additional research is required to develop a well-performing predictive model. Both the full and reduced models were dominated by abnormal urinalysis as a predictor for SBI. We identified additional historical, laboratory, and examination findings that may help identify infants with SBI to serve as candidate predictors for larger multicenter research.

Several laboratory biomarkers are likely useful in predicting SBI among hypothermic infants. Our model was primarily driven by abnormal urinalysis as a predictor of SBI, consistent with our finding that a preponderance of SBIs within this sample were UTI.²⁷  Although leukocytosis, elevated ANC, and thrombocytosis were also associated with SBI, our prior work suggests that these variables have poor accuracy when used in isolation.²¹  Our findings support using these predictors in future studies, which may be helpful in more robust risk models, particularly when using outcomes of IBI. Blood biomarkers have similarly been demonstrated to be helpful in febrile infant workup in conjunction with other laboratory indices.^15,28,29  We could not evaluate some laboratory markers, including inflammatory markers, glucose, total bilirubin, aspartate aminotransferase, alanine transaminase, and albumin because of a high proportion of missingness.

Some physical examination findings appear to also have predictive value in identifying SBIs. In the multivariable analysis of the full cohort, infants with a temperature ≤ 36.0°C had statistically higher odds of SBI. This finding was also noted in our sensitivity analysis of well-appearing infants. Using minimum temperature as a sole predictor has limited value for SBI prediction, as suggested by a recent analysis of this study sample.¹²  The severity of hypothermia may nevertheless be useful within a multivariable model. Other physical examination findings were less reliable as predictors. Abnormal examination, for example, was not associated with SBI. Although infants with hypothermia are frequently ill-appearing,³  this does not necessarily indicate the presence of bacterial infection. Unlike our study, some single-center studies report ill-appearance as a risk factor for SBI.^1,2  However, neither structured clinical scores nor unstructured clinical suspicion have been found to be reliable in identifying febrile infants with SBI.³⁰  Additionally, 1 recent single-center study found no cases of SBI among 212 well-appearing hypothermic infants referred from primary care clinics over an 8.5-year period.³¹  A lack of consensus definition for ill-appearance may be responsible for varying results between studies.

We identified historical factors, such as younger age and prematurity, that confer a lower risk of SBI. One single-center study among 360 infants found infants age 15 to 28 days were >4 times more likely (9.3% vs 2.2%) to have SBI than their younger counterparts.⁵  Wood et al similarly reported that infants ≤ 7 days of age had decreased rates of SBI as compared with infants > 7 days old: 2 of 53 (3.7%) vs 3 of 10 (30%), respectively.¹  Prior research has demonstrated an association between younger age and more extensive testing among febrile infants.¹⁵  Furthermore, single-center studies report prematurity as a potential risk factor for SBI.^1,5  Although the risk of sepsis increases with decreasing gestational age,³²  premature infants are also more prone to environmental hypothermia because of high body surface area to body mass ratio.³³  Further research is needed to better delineate prematurity as a risk factor or predictor of SBI.

Historical factors suggesting nonbacterial etiologies were less suggestive of SBI, including a history of respiratory distress and hypothermia reported at home or clinic (suggestive of environmental exposure).³¹  However, respiratory virus positivity by testing was not protective against SBI in our model, possibly because viral testing was not obtained in all cases. Importantly, a history of respiratory distress was negatively associated with SBI in the full cohort and in the sensitivity analysis among afebrile infants. Further research is required to delineate the association between viral positivity and viral symptoms with SBI in hypothermic infants.

The proportion of infants with SBI (1.7%) and IBI (0.5%) in the overall sample is consistent with prior studies on this population.^1–3,5  As febrile infants are more likely to have SBI than hypothermic infants,³⁴  applying febrile infant guidelines to hypothermic infants may result in overutilization.^15,35  Additional data to risk stratify hypothermic infants are required to direct guideline development. A complicating factor is the broad differential diagnosis for infants presenting with hypothermia, ranging from benign self-limited conditions to life-threatening illnesses.^3,5,36  Clinicians seeking to create practice pathways for hypothermic infants should consider alternative etiologies beyond SBI. In our sensitivity analysis among hypothermic infants who were afebrile at home and in the ED, we found that positive urinalysis and abnormal laboratory parameters (leukocytosis and elevated ANC) were positively associated with SBI. These factors may be useful in the risk stratification of hypothermic young infants in the absence of clear guidelines.

Our work suggests that further research is required to accurately identify IBI among hypothermic infants. This analysis was limited by the number of infants with IBI and the preponderance of UTIs overall, leading to a high number of misclassified IBIs. In this subset, leukocytosis was the only variable shown to be predictive of IBI in univariable analysis. Although WBC is associated with bacterial infections in febrile infants, its discrimination is poor.²⁸  Interestingly, all 4 cases of bacterial meningitis were found in the subgroup of afebrile infants. Although difficult to draw conclusions from this small number, it is an important reminder that isolated hypothermia can be a presenting sign of bacterial meningitis. Further research describing factors associated with IBI in hypothermic infants using larger datasets is required.

Our findings are subject to limitations, including its retrospective design and limited number of infants with SBIs. Our findings may lack generalizability since all patients were from academic children’s hospitals. As only infants with blood cultures were included, some infants with SBI may have been missed. Maternal historical factors were difficult to abstract and were frequently missing. Therefore, we could not describe maternal factors’ influence on SBI risk. Additionally, using abnormal urinalysis as a risk factor for SBI may be problematic because of its incorporation within the definition of UTI; this limitation extends to similar studies among febrile infants.¹³  Although future studies should focus on the clinically more important outcome of IBI to avoid this limitation, larger datasets are required to develop a robust model. Our sample was selected by identifying infants with a rectal temperature <36.5°C, including infants who were subsequently found to be febrile. Understandably, the presence of fever (either historically or reported) showed a strong association with SBI in the univariable analysis. Febrile and hypothermic infants may exhibit a similar bioresponse to bacterial infections,³⁷  but studies dedicated to infants with isolated hypothermia will be needed. The number of infants with isolated hypothermia (without fever) in our sample was smaller. Lastly, some hypothermic infants may have been missed if sites did not obtain a rectal temperature, and ED temperatures varied in measurement method which may have affected our interpretation of temperature data. Despite these limitations, our study provides important preliminary data describing risk factors for SBI among hypothermic infants.

Hypothermic infants presenting to the ED are at risk for SBI, though less so than their febrile infant counterparts. Historical, examination, and laboratory data are potential variables to be used in a risk prediction model to risk-stratify young hypothermic infants with concern for SBI. Although our model was a strong fit for SBI prediction, high rates of misclassification, particularly of IBI, suggest a need for additional research to safely risk-stratify this sample.

We thank Norma-Jean Simon, MPH, MPA (Smith Child Health Outcomes, Research, and Evaluation Center, Ann and Robert H. Lurie Children’s Hospital of Chicago), Chiu-Mei Chen, MS (University of Michigan), Elaine James, DNP, MSN, RN-BC (WakeMed Health and Hospitals), Kelly Huynh, MStat (University of Utah) for assistance with data acquisition.: