Predictive Model for Serious Bacterial Infections Among Infants Younger Than 3 Months of Age

DISCUSSION

Febrile infants are commonly evaluated by pediatricians and emergency physicians.¹⁰ For those infants who appear ill or who have an evident focal infection, management is straightforward. The vast majority of infants, however, appear relatively well with nonspecific symptoms and no localized source of infection. Because of the relatively high prevalence of occult infections, many of these infants undergo invasive testing to include cultures of the CSF, urine, and blood. Many of the infants are hospitalized and given empiric antibiotic therapy pending culture results. Outpatient management of a select minority of infants, who are considered to be low risk for SBI by published criteria, has become more popular over the last decade, but the selection process remains controversial.^7–9,11 Although management guidelines (based on meta-analyses of available data) have been offered in the medical literature,¹² the management of young febrile infants continues to be debated.^2,,313–17

In review of the literature regarding febrile infants, the available data are further complicated by differing and arbitrarily determined subpopulations being studied (0–3 months, 1–2 months, 1–3 months), varying temperature limits to define fever, different definitions of fever without source and SBI, and variability of testing and the criteria used to define high- versus low-risk infants. Despite the lack of universal acceptance of any one strategy, previous publications have consistently noted the inability to discriminate those with significant bacterial disease from those with simple viral illness based on observation scales or simple parameters such as height of fever or WBC count.^1,,818–28 Without such simple, reliable, discriminating tests, current strategies involve complete sepsis evaluations with outpatient treatment of the lower risk patients with⁸ or without antibiotics.^7,,9 Three common strategies for managing febrile infants without a fever source are: 1) the Philadelphia protocol for infants 29–60 days old,⁷ 2) the Rochester criteria (≤60 days old),⁹ and 3) the Boston criteria (28–89 days old).⁸ Each has its own definitions of the high-risk and low-risk infants based on a combination of factors including history, physical examination, and laboratory parameters (Table 3). In the first 2 strategies, the lower risk patients are selected for outpatient therapy without antibiotics, whereas the Boston strategy treats all patients with empiric antibiotics but selects a smaller high-risk population for hospitalization. The Philadelphia protocol and Rochester criteria have been promoted by their high NPV (ability to remove patients with SBI from the low-risk group)—99.7% and 98.9%, respectively. However, each suffers from a relatively low positive predictive value—14% and 12%, respectively—attributable to the large numbers of patients considered higher risk and therefore hospitalized for antibiotics. The Boston criteria has been argued to be the more cost-effective strategy (despite treating all with antibiotics) because fewer patients require admission.²⁹

For the current study, we wanted to develop a model to predict infants at high risk for SBI. Unlike the previously discussed criteria that had been preselected, we allowed the data to determine which clinical variables (as well as the corresponding cutoffs for continuous variables) were used. The order and use of variables are derived by the software without any bias other than the choice of potential predictors. Additionally, decision-tree analysis is easy to understand and apply as compared with other statistical methods such as logistic regression. The predictive model sequentially used four common clinical parameters: UA, WBC count, temperature, and age. With such a simple model, the NPV was 98.3% for SBI and 99.6% for bacteremia or meningitis. In a comparison to the Rochester criteria, the recursive partitioning model had a comparable enrichment of the higher versus lower risk group (Fig 3). The value of selectively enriching the high-risk group versus the low-risk group (Rochester Criteria) can be seen in the proportion of high- versus low-risk patients in each model; if the intervention was to admit only higher risk patients, the recursive partitioning model would admit only 28% versus 53% (Rochester criteria).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Comparison of recursive partitioning model to Rochester criteria.⁹

Despite the excellent NPV of our current model (similar to the Philadelphia and Rochester criteria), the limitations of any management strategy for this population are evident when looking at the SBI patients who were misclassified into the lower risk group: 1.7% (66/3814) of these patients had SBI (51 UA-negative UTI, 14 bacteremia, 1 meningitis). The relatively low sensitivity of the standard UA (dipstick plus microscopy) in this study (80%) is comparable to previous reports.³⁰ Despite the predominance of UTIs among the misclassifications, the relative frequency of misclassified patients with bacteremia and meningitis must be more seriously considered when evaluating any management strategy. Therefore, despite a comparable performance of our simple model to more complex management schemes, the significant misclassifications of SBI patients into a lower risk group suggest that an outpatient strategy using empiric therapy may be justified.

A true, direct comparison to the Philadelphia and Rochester criteria cannot be performed because each applies to specific age groups (and not all febrile infants <3 months old) and certain key data elements that are used by the Philadelphia criteria (appearance of patient, band-neutrophil ratio, urine Gram stain, CSF Gram stain, results of chest radiograph and stool smears if obtained) and the Rochester criteria (patient appearance, absolute band count, results of stool smear if obtained) are not available in our dataset. Although some patients that were misclassified by the recursive partitioning model would be properly classified by the other strategies, other patients that were correctly classified by the recursive partitioning model may be misclassified by these other strategies. In our dataset of infants <3 months with temperature ≥38.0°C, the Philadelphia criteria (using WBC count <15 000/mm³, UA ≤10 WBCs/hpf, CSF <8 WBCs/mm³) would misclassify 69 patients (16 bacteremia, 2 meningitis, 51 UTI) and the Rochester criteria (using WBC count >5000 and <15 000/mm³, UA ≤10 WBCs/hpf) would misclassify 77 patients (14 bacteremia, 6 meningitis, and 57 UTI)—the value of this analysis is only to compare the recursive partitioning model to other criterion (using our data) and not to test the other strategies. Interestingly, the Philadelphia protocol is intended for use in infants 29 to 60 days old with temperature ≥38.2°C; in this age subgroup alone, we had 1175 infants with temperatures of 38.0 or 38.1°C who would not require any diagnostic testing under the Philadelphia protocol—47 (4%) of these infants had SBI (6 bacteremia, 1 meningitis, 40 UTI), therefore the 38.0°C threshold for diagnostic testing seems justified in this age group. Additionally, both the Rochester criteria and Philadelphia protocol are directed at infants <2 months old; in our study, 7.1% (125/1753) of infants 2 to 3 months old had SBI (including 3 patients with meningitis and 14 with bacteremia).

This study and its methodology have certain limitations. By using a limited definition of SBI to include only meningitis, bacteremia, and UTI, we underestimated the prevalence of SBI in our population. Not all patients had all cultures performed—again underestimating the prevalence of SBI and potentially affecting the development of the model. We were not able to assess the value of some other predictors used by other strategies, such as Gram stains of the urine, which could easily diminish the number of misclassifications.³¹ And although a positive urine culture is the gold standard for diagnosing a UTI, it is worth mentioning that some patients with a positive urine culture and a negative UA could simply have asymptomatic bacteriuria or, theoretically, contamination. Furthermore, the appearance of the patient was not prospectively determined—ideally, ill-appearing children and those with obvious sources of focal bacterial infection would be removed before development of any management strategy (aimed at the well-appearing febrile child without an identifiable source of infection). Finally, the past medical history of the patient—including perinatal history, underlying conditions, and infectious exposures were not considered in our model but would be important factors in practice.