Discussion
This study identified that temperature 1-hr after initial acetaminophen dosing was significantly higher in individuals with culture confirmed bacteremia. We demonstrated that temperature 1-hr after initial acetaminophen had significant value in predicting bacteremia using both a binary logistic regression and a CART machine learning model, with AUC-ROC comparable to other prospectively validated CDRs.16,19-22 Similar to prior reports, we observed that presenting temperature ≥ 39.0°C, evidence of focal infection and hypotension were also all predictive of bacteremia5,8,9,15. However, when examining the relative importance of input variables in our CART model, temperature 1-hr after acetaminophen was far and away the most important predictor of bacteremia when compared to pre-acetaminophen temperature ≥ 39.0°C, evidence of focal infection, and hypotension. Generation of an optimal decision tree using CART analysis also demonstrated that 85.7% of patients with temperature ≤ 37.3°C (99.2°F) 1-hr after acetaminophen had negative blood cultures. Conversely, 80% of patients with temperature > 39.2°C (102.5°F) 1-hr after acetaminophen were culture positive. These results suggest there may be some utility in using post-acetaminophen temperatures at both extremes of the spectrum to aid current CDR’s in identifying patients at low- and high-risk of bacteremia. This may be especially useful in assessing patients with a mix of low- and high-risk features (ex: high degree of neutropenia with low-risk category of malignancy), in which prompt and accurate determination of risk can be difficult.
Temperature has primarily been used as a binary variable (ex. Temperature ≥ or < 39.0°C) in CDRs for FN risk stratification, with few if any studies investigating the relationship between fever response to antipyretics and bacteremia. Alali et al. described the creation of a random forest model for predicting BSI and need for transfer to an intensive care unit (TIC), in which the authors detail that treating temperature as a continuous variable and including peak temperature as an input variable increased model performance.16 The authors also identified max temperature to be the most important predictor of bacteremia and the second most important predictor of TIC (behind only hypotension).16 While our study did not demonstrate a significant difference in Tmax between patients with bacteremia and culture negative individuals, the concept of using temperature to predict bacteremia as well as the relative importance of temperature as a predictor of bacteremia are congruent with the results of this study.
Another study by Haeusler et al. utilized prospective data from the Australian-PICNICC study to investigate the association between 10 commonly utilized CDR variables and the presence of bacterial infection.22 Using both univariate and multivariate analysis of 858 episodes of FN, the authors demonstrated that increasing temperature was significantly associated with bacterial infection and was one of only 3 variables that maintained significant predictability for bacteremia after multivariate analysis (the other two being decreasing platelet count and appearing clinically unwell).22 These results again highlight the important relationship between temperature and bacteremia in FN and lend credence to the idea that fever response 1-hr after acetaminophen is capable of lending valuable insight into which patients are likely to have bacteremia and should not be categorized as low risk.
Many contemporary CDRs utilize a host of patient or disease related factors to help risk stratify patients presenting with FN (specific type of leukemia/lymphoma, presence of relapse with marrow involvement, chemotherapy within 7 days of FN presentation, etc.).6-8,10 To control for these factors, we split patients into groups based on their category of malignancy (leukemia/lymphoma, solid tumor or HSCT), and matched them based not only on this category but also on the degree of neutropenia, using this as a proxy for relative immunosuppression. Another commonly utilized patient factor for risk stratification is presence of a central venous catheter (CVC). CVCs are present almost ubiquitously across this institution for patients undergoing active treatment and thus was deemed largely unhelpful and non-discriminatory for the purposes of this investigation. Rondinelli et al. also cited age as a patient factor that could increase the risk of serious infectious complications, assigning increased risk to patients ≤ 5 years old. In our study, by matching cohorts-controls based on age and ensuring all cohort patients ≤ 5 years old were also matched with controls ≤ 5 years old, we made the response to acetaminophen variable independent of age.
Several laboratory results and biomarkers are also commonly utilized in FN CDRs. Inflammatory markers such as C-reactive protein (CRP), procalcitonin (PCT), interleukin-6 (IL-6) or interleukin-8 (IL-8) have all been touted as helpful means of FN risk stratification.6,10,23-25 The use of CRP, PCT, IL-6 and IL-8 as adjunct to fever response to acetaminophen were investigated for this study; however, at our institution like many others, it is not routine practice to obtain these inflammatory markers on presentation and thus the retrospective data available for our patient population was too sparse to provide any meaningful insights. AMC is another laboratory variable seen in several CDRs. AMC cutoffs for determination of high-risk patients range between < 100 cells/µL to < 155 cells/µL depending on the CDR.5,11,26 In this study we did not see a significant relationship between AMC and bacteremia at a cutoff of < 100 cells/µL. One prospective validation study of several CDRs found that by altering the cutoff parameters and recalibrating several of these CDRs they were able to raise the overall AUC-ROC and low-risk yield for the risk stratification models.22 These results were obtained in part by lowering the AMC cutoff from < 100 or < 155 cells/µL to < 15 cells/µL, suggesting that the lack of significance between AMC and bacteremia seen in our study may be due in part to questions regarding the optimal AMC threshold.
Our study has several limitations. As an analysis of a single tertiary academic medical center, our results may not necessarily be generalizable to other institutions. Our study is also limited by its retrospective nature which hindered the accurate collection and incorporation of clinical and laboratory variables (ex. GI symptoms, chills/rigors, mucositis, evidence of focal infection, CRP, PCT, IL-6, IL-8) into our models; however, all single-center retrospective analyses suffer similar drawbacks, making the need for prospective analysis and multicenter validation critical.