C-reactive protein for late-onset sepsis diagnosis in very low birth weight infants

This retrospective study was performed at the CHU de Quebec – Université Laval NICU (Quebec, Canada), between March 2008 and April 2013. The study was approved by the Institutional Research and Ethics Board. The dataset analyzed for the current study is available from the corresponding author upon reasonable request. All neonates with a birth weight < 1500 g and more than three days old at initial late-onset sepsis evaluation were included.

The hospital clinical database Med-Echo, a national validated medico-administrative database was used to collect patient demographics [21]. The local infectious disease database TDR was used to collect dates of blood cultures. Individual patient laboratory and clinical data were collected using electronic medical charts. Time of initial evaluation (T0) was defined as the of the initial blood culture. CBC and CRP values were collected at T0, at 16–24 h (T24) and 40–48 h (T48) after.

CRP levels were determined using Vitros CRP slide method (Vitros 250 Chemistry System, Ortho-Clinical Diagnostic, Johnson and Johnson). The CRP was considered positive if the laboratory value was ≥10 mg/L. The CBC was considered abnormal if any of the following was present: white blood cell count (WBC) < 5000/mm³, I/T ratio > 0.10 or platelet count < 100,000/uL.

Continuous variables with normal distributions are presented with the mean and standard deviation, while continuous variables with non-normal distributions are presented with the median and interquartile range (IQR). Qualitative variables are presented with frequency and percentage. Analysis of variance, estimated with generalized estimated equations (GEE), was used to test for differences in median CRP results at T0, T24 and T48 between the different bacterial pathogens. GEE were used to calculate sensitivity, specificity, positive and negative predictive values with 95% confidence intervals (CI) for predicting late-onset sepsis for the CRP and CBC at T0, T24 and T48 as well as for different combinations of these tests. Logistic regressions were used to compare sensitivity and specificity between the CRP, the CBC and different combinations. Multiple comparisons were corrected using Bonferroni method. Patients with rapidly resolving clinical symptoms with a normal CBC and CRP at T0 might not have had repeat testing done at T24 and T48 and patients with abnormal results (CBC or CRP) at T0 may not have had repeated tests since the early markers were already positive. Both situations could lead to a selection bias by analyzing the available values at those time points because using only available data may change the prevalence of the disease and affect the diagnostic accuracy of the blood tests [22]. Consequently, last observation carried forward (LOCF) imputation technique was used for missing values when data at previous times was available to limit a potential selection bias.

During the 5-year period, 1090 blood cultures were performed in 416 eligible VLBW infants. A total of 590 distinct late-onset sepsis evaluations met the inclusion criteria. Of the 500 excluded blood cultures, 481 represented repeated blood cultures done within 14 days of the initial late-onset sepsis evaluation, 9 were probable contaminants and 10 had no available associated CRP data at all three time points. The demographic characteristics of the patients included are detailed in Table 1. In total, 162 (27%) evaluations were culture proven late-onset sepsis, all had a least one positive blood culture. CoNS were isolated in 83% of the episodes of infection, and the remainder were caused by other gram-positive bacteria (9%), gram-negative bacteria (7%) and fungi (1%). Among the 162 blood culture-proven late-onset sepsis episodes, 3 had meningitis (1 fungal and 2 bacterial) and 6 had a urinary tract infection. There were no cases of meningitis with a negative blood culture.

After LOCF imputation, there were 575, 583 and 586 available CRP values at T0, T24 and T48 respectively. CRP peaked at 24 h irrespective of the causative pathogen. At T24, the median CRP values were 38 mg/L for gram-positive bacterial infections other than CoNS, 40 mg/L for CoNS infections and 90 mg/L for gram-negative bacterial infections (Table 2). At T0, T24 and T48, all comparisons of CRP values between pathogens were not statistically significantly different (all P values > 0.80). There were 116 (48%) false positive CRP tests at T24 (CRP > 10 and no late-onset sepsis): 14 were treated for necrotizing enterocolitis, 8 had tests taken when the infant was in a postoperative period and 38 had suspected ventilator-associated pneumonia. At T24, there were 25 (8%) false negative CRP tests (CRP ≤10 with late-onset sepsis) for which the causative organisms are listed in Table 3.

The sensitivity, specificity, positive and negative predictive values of the CRP, CBC and their combinations were calculated at each time point and compared (Table 4). At initial sepsis work-up (T0), combining the CBC and CRP offered the highest sensitivity for late-onset sepsis diagnosis (65%) which was statistically superior to CRP (49%, p < 0.001) and CBC (49%, p < 0.001) alone. At T24, the sensitivity of the CRP increased, and was not statistically significantly different than when combined with the CBC (84% vs 87%, p = 0.36). At T24, the sensitivities of the individual components of the CBC were 50% for the I/T ratio, 4% for leukopenia and 12% for thrombocytopenia. Compared to its performance at T24, the sensitivity of the CRP at T48 decreased (84% vs 73%, p = 0.08). At T48, the sensitivity of the combined CBC and CRP was similar to the individual CRP (76% vs 73%, p = 0.35). The sensitivity of the CBC at T48 significantly decreased compared to T24 (21% vs 59%, p < 0.001).

Table 5 presents multiple test combinations at different times. The maximum sensitivity and negative predictive values were 88% and 93% respectively, and could be obtained by performing only a CBC at T0 with a CRP at T24. Also, the sensitivity obtained with three consecutive CRP measurements at T0, T24 and T48 was 86% which was not superior to the sensitivity of a single CRP at T24 (p = 0.93).

ROC curves and the corresponding AUC of different tests are presented in Fig. 1. The AUC of a single CRP measured at T24 was 0.82 and was not statistically significantly different than the variations in CRP, white blood cell count and I/T ratio (all P values > 0.75).

This study focused on the use of CRP as an adjunctive diagnostic tool in the evaluation of VLBW infants with suspected late-onset sepsis. The sensitivity of the CRP at initial evaluation (T0) was low (49%), which correlates with previous studies in other neonatal populations [12, 15]. Nevertheless, at 24 h, CRP had a better sensitivity of 84% for late-onset sepsis and a high negative predictive value of 92%. This is similar to previously published results in in cohorts of infants born at different gestational ages [12, 15]. Serial measurement of CRP at T0, T24 and T48 was associated with a 93% negative predictive value, which is lower than what other prospective studies have reported (98%) [16]. However, those studies included smaller numbers of VLBW infants and had a higher prevalence of gram-negative bacterial sepsis.

We report a high rate of CoNS sepsis compared to previous studies on CRP [14, 16, 20]. However, these rates are comparable to the incidence of CoNS infections in the Canadian Neonatal Network (> 70% of late-onset sepsis are caused by CoNS) [18]. This is likely attributable to variations in local epidemiology combined with care practices. Indeed, the rate of fungal infections was also very low in our cohort and similar to the Canadian average (incidence < 2% in VLBW infants) [18]. The fact that we did not mandate two cultures for the diagnosis of sepsis may have contributed to a higher prevalence of contaminants. This may have increased the false positive rate in our cohort, but would have little effect on the negative predictive value of CRP.

Previous studies concluded that CoNS might induce a less sustained inflammatory response than gram-negative bacterial sepsis [13, 14]. However, our findings do not suggest a statistically significant difference in peak CRP’s values in infants with CoNS compared to other pathogens, however this is not statistically significant given the small number of gram-negative infections. Likewise, 46% of the CoNS sepsis included in our cohort were associated with a CRP value of at least 50 mg/L suggesting that the inflammatory potential of these organisms is present.

We found that a CBC obtained at T24 has a low sensitivity (59%) for the diagnosis of late-onset sepsis in VLBW infants. The I/T ratio was the main contributor to the diagnostic accuracy of the CBC at T24 (sensitivity 50%). This is in keeping to what has been reported in more mature newborns. Indeed, Hornik et al. found that the sensitivity of the CBC in infants born < 34 weeks gestational age was 55% [7]. However, they used a different cut-off for the I/T ratio (> 0.2). They noted a higher I/T ratio in gram-negative infections. The low incidence of gram-negative infections in our cohort might explain why a lower threshold yielded similar results. Moreover, as a late-onset physiologic neutropenia frequently occurs in VLBW infants, their I/T ratio might be less accurate [23]. Consequently, using a lower positivity threshold may be appropriate in order to increase its sensitivity as a screening tool in this specific population. A platelet count < 100,000/uL at T24 had a low sensitivity (12%) for predicting late-onset sepsis. This is similar to previous studies that found the platelet count has a low discriminative performance (AUC 0.60) for diagnosing sepsis in preterm infants [7]. The variation in platelet count based on gestational age and postnatal age may require the use of age specific cut-offs [24]. Also, platelet parameters like mean platelet volume and platelet distribution width may increase diagnostic yield, although these parameters were not collected in the present study [25].

The combination of the CBC at T0 and CRP at T24 had the highest negative predictive value (93%) with a minimal number of tests. Considering every blood drawn represents a significant volume loss for premature neonates and carries an inherent infection risk, the use of a less invasive, but still accurate sepsis work-up strategy is worth considering. Since the majority of clinicians include a CBC in the initial workup of suspected late-onset sepsis [7], it is unlikely that an approach solely based on serial CRP measure would be adopted. Also, two CRP values drawn at T0 and T24 had a non-superior negative predictive value than the combined CBC at T0 and CRP at T24. Further, there were no additional benefit to repeat any blood tests at T48. Lastly, if diagnostic accuracy at the time of initial evaluation (T0) is a priority, then the combination of the CBC and CRP at T0 allows the highest sensitivity and might assist in clinical decision-making.

An important benefit of using reliable laboratory markers is to help in the decision to discontinue empirical antimicrobial. Recent studies reported a 97% negative predictive value of blood culture at 36 h [26, 27]. The high negative predictive value of negative blood cultures at 36 h, combined with the 93% negative predictive value of the two-step approach described above (CBC at T0 and CRP at T24) could be used together to support discontinuation of antimicrobials at that time rather than waiting the traditional 48-h time point. This, in turn, would help mitigate drug-associated adverse events, decrease the likelihood of selection of resistant organisms and Candida sp. [26], necrotizing enterocolitis and potential hearing impairment [28], in addition to reducing healthcare costs [6]. Additionally, discontinuing antibiotics after 36 h reduces blood draws required for antibiotic dosing when using aminoglycosides and/or vancomycin are used as empiric therapy.

The analysis of ROC curves associated with the absolute change from T0 to T24 of the CRP, the WBC count and the I/T ratio has clinical implications. For example, it improves the WBC count relevance compared to a single value at T0. These results reinforce the importance of a marker’s kinetic rather than its absolute value at a specific timing. However, when trying to minimize the amount of blood tests, one should consider that the AUC of the change in white blood cell count from T0 to T24 was not significantly different than a single CRP done at T24. Consequently, repeating a CBC to monitor the white blood cell count variation is not warranted. Furthermore, as there is no current positivity cut-off for the absolute increase in white blood cell count, its interpretation is subject to variability.

An important strength of this study is the single cohort of VLBW infants. The retrospective design and inclusion of all newborns with a late-onset sepsis evaluation may have led to the inclusion of patients with other conditions associated with an increased CRP. Indeed, among infants with false positive CRP results at T24, there were 8 postoperative infants and 14 who were treated for necrotizing enterocolitis, two conditions known to increase CRP values [2, 10, 29]. Also, infants with a diagnosis of ventilator-associated pneumonia and elevated CRP were not considered as having a late-onset sepsis since the objective of the study was to assess diagnostic accuracy in predicting culture-proven bloodstream infections or meningitis. Also, there were no standardized criteria for diagnosis of ventilator-associated pneumonia in VLBW infants on the NICU during the study period. We did not collect data on vaccination and intraventricular hemorrhage which might also have increased false positive rates [10]. All these conditions, by increasing the CRP false-positive results may have contributed to underestimating its specificity. However, the inclusion of these patients would have had little effect on the negative predictive value. Finally even though missing data were imputed using LOCF method, their proportion varied between 5% and 33% for the different lab tests and time points. To ensure that this did not bias the results, sensitivity analysis without imputation were carried and showed identical conclusions.