Imagine a 9-week-old infant presents to the ED. They appear well, they take a feed, and they coo softly as you examine them, giving you that “ahh” factor. However, their temperature is 38.6ºC.
What should you do? Obtain labs? Admit? After all, young babies with fever can be deceptive. You don’t want to miss serious infections, do you? What’s this? A trainee hands you a new paper addressing your very issue. Let’s dive in.
The AAP guideline for treating febrile infants aged 8-60 days was published in 2021 (2). This guideline has since assisted clinicians in unpicking the often-tricky management of the well-appearing febrile infant. However, we are often still vigilant about an increased risk of serious infections for babies up to 3 months of age. How do we, as clinicians, ensure that infants between 60 and 90 days old don’t fall through the cracks?
Fortunately, the PECARN team have set out to answer this very question.
Their work proposes not one, but two different internally validated prediction rules that might assist you the next time you’re faced with the boiling but babbling baby. So, let’s dive in and see what they have found.
What did they do?
This was a retrospective cohort analysis of the PECARN registry spanning January 1, 2012, through April 30, 2024.
There were 4,952 well-appearing febrile infants aged 2-3 months who underwent blood culture and urinalysis.
Of these 100 children (~2%) had invasive bacterial infection (IBI). This was defined as either bacteraemia or bacterial meningitis.
What were the results?
They built a prediction rule for febrile infants 2-3 months old with statistical modelling, which demonstrated that having either:-
- Positive urinalysis
- Temperature of > 38.9°C
This yielded a sensitivity of 86% and a specificity of 58.9% for invasive bacterial infections. This led to a negative predictive value (NPV) of 99.5 (99.2, 99.7) and a negative likelihood ratio of 0.24 (0.15, 0.39) for this prediction rule.
They also conducted a secondary analysis of individuals with measured procalcitonin and absolute neutrophil counts. In this smaller cohort, 27 children had IBI s, with 1207 having these tests performed. From this, they found that children with both:
- Procalcitonin (PCT) > 0.24 ng/mL
- Absolute neutrophil count >10,710 cells/mm³
This had a sensitivity of 100% and a specificity of 65.8% for IBI. This yielded a negative predictive value (NPV) of 100.0% (99.5-100.0).
Both prediction rules were designed to minimise the number of missed cases by imposing a cost on the algorithm for producing false negatives. This partly explains the higher sensitivities and relatively lower specificities present in the models, though the specificity is on par with the PCT-based PECARN prediction rule for febrile infants <=60 days old.
They also used a 10-fold cross-validation to test these models. This is where the data is split into 10 blocks, with nine blocks used to train the model and the remaining block used to test it. This is repeated 10 times, each time with a new test block, to identify the optimal true cut-off values and assess the model’s performance.
Cross-validation showed that the models performed well when evaluated on data they hadn’t encountered during training.
Both prediction rules were developed with a strong emphasis on minimising missed cases. To achieve this, the models penalised false negatives—cases where a invasive bacterial infection (IBI) might be missed. This approach helps explain the observed high sensitivities, albeit at the cost of slightly lower specificities. However, the specificity is still comparable to that of the PCT-based PECARN rule for febrile infants aged ≤60 days.
To assess the models, the researchers used 10-fold cross-validation—a robust method for evaluating predictive performance. In this approach, the dataset is divided into 10 parts (or ‘folds’). The model is trained on nine of these folds and tested on the remaining one. This process is repeated 10 times, with each fold serving as the test set once. This technique enables researchers to identify optimal cutoff values and evaluate the model’s performance on unseen data.
The results showed that both models maintained good performance when tested on data they hadn’t been trained on, supporting their potential clinical utility.
What did it add?
These prediction rules are statistically derived from a large dataset, lending them robustness and reliability. They offer a useful tool for identifying lower-risk febrile infants aged 60–90 days—a group that has historically been underrepresented in research. Both models performed well in detecting invasive bacterial infections (IBIs). However, it’s worth noting that the simpler model—based solely on urinalysis and temperature—missed 14 cases of IBI.
In contrast, the model incorporating procalcitonin and absolute neutrophil count (ANC) was more sensitive. Among infants who underwent these blood tests, the model successfully identified all cases of invasive bacterial infection.
These are the first statistically derived models from such a large cohort focusing on this specific age group. They provide valuable, evidence-based data to support clinical decision-making in an area where guidance has previously been limited.
What were the limitations?
As with most clinical tools proposed for identifying IBIs in young children, the specificities reported are low. If this rule were to be used by itself to make treatment decisions, then it is likely that many children will receive treatment, and some will be hospitalised without having IBIs.
However, evidence shows that clinicians might prefer to err on the side of caution in such cases. So, it would be a reasonable trade-off to catch those who may become seriously unwell.
It’s also worth noting that the ANC cutoff is relatively high in this paper. This is likely to be due to the presence of pneumococcal bacteremia in this age group, which leads to higher ABC and ANC counts. It may also reflect selection bias, as infants who have undergone blood tests may represent sicker patients. If this is the case, however, it would imply that there is some part of the clinical gestalt not studied in this model, meaning that these children are more likely to have blood drawn.
What did the authors conclude, and what does it mean for current practice?
These rules show some promise in helping clinicians make decisions about the febrile but happy infant between 60 and 90 days. If you draw blood, using the absolute neutrophil count and procalcitonin tool may be particularly useful given their strong negative predictive values. These tools, however, must be externally validated through prospective studies before being used with confidence in practice. Ultimately, both prediction rules show promise in addressing the conundrum of the toasty two- to three-month-old.
CASP Checklist for cohort studies
Does this address a clearly focused issue? Yes.
Was the cohort recruited in an acceptable way? Yes. Retrospective cohort through a large registry, giving a good sample of high-quality data with clear outcomes.
Was the exposure accurately measured to minimise bias?
Yes. A natural language algorithm was developed to identify eligible participants, representative of all considered. Excluded repeat visits to avoid confounding and only considered the index visit, which is appropriate. The PECARN network, with such a high-resource population, may not be generalisable to poorer-resource settings.
Was the outcome accurately measured to minimise bias? Yes. Clear and sensible outcomes consistent with previous evidence base in the field.
Have the authors identified all-important confounding factors? Objective measurements are well-defined. Measurements reflective of practice and desirable endpoints. An NLP algorithm was developed, with a PPV of 86%. Some cases may be missed, but given the sheer volume of patients, this is an acceptable method. Measurements are correct across groups. No comment on blinding; unclear how this would have affected the results, given that machine learning is used to generate decision trees and therefore is unlikely to be affected by a lack of blinding in the same way as humans.
Was the follow-up of subjects complete and accurate? No. Retrospective, so only pertinent results are available.
Do you believe the results? Yes. Standard internal validation methods were used and well executed. It may be biased against decision trees because sicker patients may be included. The results are likely reliable given the methodology employed. Difficult to know if resource-poor areas would have the same results or areas with differing epidemiology of communicable disease.
Can the results be applied to a local population? Yes. A cohort study across such a large population is an excellent way to generate such models; they should be tested in further work to ensure generalisability to this population and others in prospective work.
Do the results fit with other evidence available? Yes.
Comment from the author, Paul Aronson
I want to first and foremost acknowledge my wonderful co-authors – Nate, Prashant, and Bob, and the PECARN EDC team, Huong, Blake, Cody, and Charlie.
The take home message I hope to get across is that when a clinician decides to obtain blood testing in a febrile infant 61 to 90 days old, based on height of temperature, clinician appearance, lack of respiratory symptoms, or other factors, we hope that our PCT-based prediction rule can provide evidence as to what tests to obtain and the cutoff values that clinicians should use for risk stratification.
Although external prospective validation is needed, this will take time and funding, and we believe our study provides the best evidence to date to help guide practice in this age group.
