Identification of factors predicting low-risk febrile neutropenia admissions in adults with acute myeloid leukemia

Febrile neutropenia (FN) is the most common complication after intensive chemotherapy for acute myeloid leukemia (AML).^1,2 The current standard approach for FN is hospital admission for IV antibiotics due to concern for high infection-related morbidity and mortality; thus FN is a major driver of hospital readmission and acute care bed utilization after therapy for AML.^2,3 Whether this approach is justified in all clinical scenarios is unknown. Furthermore, in at least 15% to 30% of FN cases, the causative infectious pathogens cannot be identified despite thorough diagnostic evaluations,⁴ implicating noninfectious causes of fever in some cases, such as disease, mucosal inflammation, medications, or transfusions, which might not warrant hospital admission.^5-8 Admitting all patients despite the potential for less severe episodes of FN or noninfectious causes may not be beneficial. Thus, the ability to identify these episodes of “lower-risk (LR)” FN upfront is appealing to expedite clinical decision-making and potentially avoid hospitalizations that contribute to the current inpatient bed shortage, increase the use of health care resources, and negatively impact patient quality of life.

Multiple studies have been done to create models and guidelines that aim to define a low-risk group of oncology patients with FN (Multinational Association of Supportive Care in Cancer [MASCC] score,⁹ clinical index of stable febrile neutropenia [CISNE] score,¹⁰ Infectious Disease Society of America guidelines,¹¹ and National Comprehensive Cancer Center guidelines¹²). These have largely been developed and validated in solid tumors, and many exclude patients with hematologic malignancies or expected severe neutropenia. These tools have led to several randomized clinical trials in solid tumors that have provided evidence supporting the feasibility and safety of ambulatory treatment for low-risk FN.⁶ When patients with AML are included, they are often automatically categorized into higher-risk (HR) groups, thus excluding them a priori from potential low-risk categorization.⁷ 

With the decline in mortality after AML therapy,^8,13 improvements in our ability to provide complex care at outpatient facilities^14,15 and/or at home (eg, via home infusion services), and the availability of oral broad-spectrum antimicrobials that can replace IV medications in certain circumstances, outpatient treatment of FN in AML may now be feasible in selected, otherwise medically stable patients, if they can be identified. We, therefore, aim in this study to (1) determine whether baseline demographic, disease-related, and clinical characteristics can predict which patients with AML and FN have a lower risk of progression to severe illness, and (2) evaluate the performance of established risk models in this population.

This is a single-center, retrospective cohort study conducted at the University of Washington/Fred Hutchison Cancer Center. Records from 2016 to 2023 were identified via our institutional AML database and supplemented by reviewing inpatient oncology admission lists in the electronic medical record (EMR). Subjects were then screened for eligibility and episodes of FN. We included adults aged ≥18 years with newly diagnosed (ND), relapsed, or refractory high-grade myeloid neoplasms (≥10% blasts in the blood/marrow) who had received intensive chemotherapy in the preceding 8 weeks and were then discharged from the hospital during their cytopenic period. At our institution, we discharge patients immediately after the completion of induction or postremission therapy as long as they can stay within 60 minutes of our center and have a caregiver.¹⁴ Intensive chemotherapy was defined as a regimen including a backbone of “7 + 3,” CPX-351, or cytarabine at a dose ≥1000 mg/m² per day, and could be given either as (re)induction or postremission (eg, “consolidation”) therapy. Prior allogeneic hematopoietic cell transplantation (HCT) was allowed if it was at least 3 months before the FN episode. FN was defined as a fever ≥38.3°C with neutrophil count <1000 cells per μL measured within the last 48 hours.¹⁶ 

Baseline demographic, clinical, and disease variables were collected including age, gender, disease status (ND or relapsed/refractory), history and date of HCT, chemotherapy regimen (categorized as “standard intensity” [“7 + 3” or CPX-351] and “high-intensity” [cytarabine doses of ≥1000 mg/m²; regimens such as fludarabine, high-dose cyarabine, G-CSF, and idarubicin or cladribine, high-dose cytarabine, G-CSF and mitoxantrone; which is the standard backbone at our institution^17,18), duration of neutropenia prior to the FN episode, blood product transfusion within the prior 24 hours of the FN episode, localizing symptoms on presentation, Eastern Cooperative Group performance status (PS) on presentation, fever variables (initial temperature, highest temperature during admission, and duration of fever), laboratory studies on presentation, vital signs (VS) on presentation and in the first 24 hours, prior infections and major comorbidities. We recorded infection identified, site of infection, type of organism, and radiologic evidence of infection. Identified organisms were classified as bacterial (gram-positive, gram-negative, or polymicrobial), viral, or fungal. A proven bacterial infection was only assumed if culture-documented. Fungal infections were categorized as proven or probable based on the European Organization for Research and Treatment of Cancer/Mycoses Study Group Working Group criteria.¹⁹ Microbiological documentation (typically via polymerase chain reaction) was required for viral infections. We recorded the need for intensive care unit (ICU) care, hospital length of stay (LOS), in-hospital death, and survival. Finally, we collected variables specifically used in other currently available risk models for both FN and ICU outcomes for each patient, including the MASCC risk index for FN, CISNE, Charlson comorbidity index (CCI),²⁰ acute physiology and chronic health evaluation (APACHE II),²¹ and quick sequential organ failure assessment (qSOFA) scores^22-24 (fully described, including components in supplemental Table 1).

A "LR" outcome after FN admission was defined as a hospital LOS of <72 hours without the need for ICU admission or inpatient death. A “HR” admission was defined if any of the above occurred. The primary objective of this study was to identify patents with AML who would have an LR outcome and to determine whether the existing FN and critical illness risk models could identify this population. The secondary objective was to identify the factors that predict the identification of an infection as the etiology of the FN episode. Patient characteristics were compared using Fisher exact test (categorical characteristics) and Wilcoxon-Mann-Whitney tests (quantitative characteristics). Univariate and multivariable (MV) logistic regression mixed-effects models (accounting for patients with >1 admission) assessed covariate associations with an LR admission and infection identification. The covariates included in the regression models were selected based on their univariate area under the receiver operating characteristic curve (AUC) values. Two-sided P values were reported; no corrections were made for multiple comparisons. The optimism-adjusted AUC was reported for MV models. Statistical analyses were performed using the R software (http://www.r-project.org). This retrospective analysis was approved by Fred Hutch’s institutional review board.

We identified 781 patients from our institutional AML database and review of oncology admissions in the EMR from May 2016 to May 2023. A total of 533 patients were excluded due to the type of chemotherapy regimen (eg, lower intensity; 50%), no FN admission (19%), admitted during use of a prior EMR not accessible for data collection (17%), FN admitted to another institution (6%), fever during chemotherapy before discharge (6%), or wrong diagnosis (3%), leaving 248 distinct patients with 397 FN admissions that met our criteria (CONSORT diagram, supplemental Figure 1). For all admissions, the median age of the patients was 61 (range, 21-87) years, 70% were ND, 30% were relapsed/refractory, 7% had prior HCT, and almost all (99%) received high-intensity chemotherapy. The baseline patient characteristics broken down by LR vs HR first admission are shown in Table 1. The median duration of neutropenia before FN was 12 (1-228) days. Forty-seven percent had transfusions within 24 hours before fever. The median time from day 1 of the most recent chemotherapy course to the first episode of FN was 15 days, and the median hospital LOS was 6 (1-56) days; in 10% of admissions, patients required ICU care, and 3.5% of patients died inpatient.

First, we focused on the initial FN admission for the 248 patients. Overall, 15% of admissions were classified as LR and 85% as HR (distribution of hospital LOS in Figure 1). Infections were identified in 26% of LR admissions and 66% of HR admissions (P < .001). The baseline/initial characteristics associated with LR admissions in univariate analysis (Table 2) were older age, higher absolute monocyte count, and lower initial heart rate. In the first 24 hours, higher systolic blood pressure (BP), lower heart rate, lower respiratory rate (RR), lower fever, and highest fever <39°C were significantly associated with LR admissions. We found no association between transfusion within 24 hours, disease status, secondary disease, prior transplant, duration of prior neutropenia, performance status (PS) at presentation, any comorbidity, prior fungal infection, most other baseline laboratory values, and LR vs HR admission. We repeated this analysis including all FN admissions for each patient (total n = 397, 14% were classified as LR admissions) and found no major differences in predictive factors for LR admission compared with the initial cohort of first admissions only.

To further elucidate the relationship between these factors and admission risk, we next created a MV logistic regression model to predict LR admissions using only initial or “presenting” variables, and then the “most abnormal” variables in the first 24 hours of admission (Table 3). The model based on only presenting variables included initial heart rate, RR, and absolute monocyte count, and had an optimism-adjusted AUC for prediction of LR admission of 0.70. The model based on the most abnormal 24-hour variables, including the highest fever, highest heart rate, highest RR, and age, had an AUC of 0.82.

Infection was identified as a source of FN in 59% of FN admissions and primarily consisted of blood stream infections (supplemental Tables 2 and 3 describe the type and site of infections identified). Baseline factors associated with the identification of infection as an etiology of FN in univariate analysis (Table 4) included male sex, relapsed/refractory disease (vs ND), lower initial absolute lymphocyte count (ALC), lower potassium, higher anion gap, and higher initial RR. In evaluating the most abnormal of each variable in the first 24 hours of admission, lower systolic BP, lower mean arterial pressure, higher heart rate, higher RR, higher fever, and fever > 39°C were all associated with identification of infection. Unlike LR admissions, age was not associated with infection identification. However, similar to illness severity, we found no association between transfusion within 24 hours, secondary disease, prior transplant, duration of prior neutropenia, PS, any comorbidity, prior fungal infection, other baseline laboratory values, and LR vs HR admission. Applying this analysis to both initial and subsequent FN admissions did not change the results.

In the MV logistic regression analysis (Table 3), the model based on only presenting variables including initial heart rate, RR, and ALC had an optimism-adjusted AUC for the prediction of infection identification of 0.63. The model based on most abnormal variables in the first 24 hours of admission included the highest fever, highest heart rate, highest RR, and lowest ALC with an AUC of 0.72.

Although age was significant in both the univariate and multivariate models for LR admission, it was in the reverse direction of the established models, with older age associated with LR admissions. Age was not associated with an infectious etiology identified as a cause for the FN. To better understand this first finding, we evaluated differences between patients aged ≤65 years (n = 175) and those aged >65 years (n = 73) and found that older patients were more likely to have secondary disease (P = .02), higher PS (P < .001), cardiac comorbidities (P < .001), and a lower initial heart rate (P = .003); and anion gap (P = .002); none of these would explain an association with LR admissions. There were no differences between these cohorts in recent transfusion, prior fungal infection, disease status, prior HCT, other baseline laboratory values or physiologic variables, or whether infectious etiology was identified (P = nonsignificant for all).

Finally, we evaluated the performance of previously developed FN risk and critical illness risk models in this AML population and compared their predictive performance using C-statistics to our own MV models predicting LR FN admission and infection identification (Table 5). The AUC for the prediction of an LR admission for all established models ranged from an AUC of 0.56 for the APACHE II model to an AUC of 0.62 for the CCI. The CCI was the worst at predicting identification of infection, with an AUC of 0.51, and the MASCC was the best, with an AUC of 0.69.

In this study, we aimed to identify the predictors of LR FN admissions in patients with AML undergoing intensive chemotherapy and to evaluate existing risk models in this domain. Using our prespecified definition of LR FN, we found that only 15% of patients with AML admitted for FN met these criteria and that an infectious etiology for the fever could be identified in only 59% of these cases. Physiologic parameters including heart rate, BP, and height of fever were best able to predict LR admissions and infection identification. Factors that predict HR outcomes in established FN models such as the MASCC score, such as comorbidities, older age, and prior fungal infection, were not predictive of HR outcomes in AML. We thus developed our own MV model based on most abnormal variables within 24 hours of admission with an AUC of 0.82 to predict LR admissions and an AUC of 0.72 for the prediction of an infectious etiology identified.

Our study specifically examined the performance of existing risk models in patients with AML, a group often excluded from initial development and validation studies for these models. The MASCC, for example, is one of the most widely used to determine the need for inpatient admission for oncology patients with FN. It has been validated for all cancer patients but excludes patients with acute leukemia undergoing induction chemotherapy or transplantation. Similarly, the CISNE score, validated primarily in patients with solid tumors, uses the absence of profound neutropenia to help stratify patients into low- or high-risk categories for complications, effectively shunting all patients with AML into the high-risk category. This study demonstrates that these existing FN risk models and general models used to predict the development of critical illness in broader populations (qSOFA, APACHE) do not perform as well in predicting LR FN admissions for patients with AML. This finding is supported by a similar study²³ in patients with AML and FN evaluating the qSOFA (which includes altered mental status, RR, and systolic BP) in predicting outcomes, which found that although the qSOFA scores identified patients with AML at risk for severe outcomes with high specificity, they had low sensitivity. This study, however, aligns with our observations that physiological parameters play a significant role in predicting which patients with AML develop more severe illnesses. Although these models have shown utility in other cancer types, our model outperformed these existing scores in predicting LR admissions, highlighting the need for disease-specific risk stratification tools in patients with AML undergoing intensive chemotherapy.

Our best-performing models were largely based on VS, underscoring the importance of physiological parameters for risk stratification in this setting. One notable finding was that the height of the fever was associated with both HR outcomes after admission and an infectious etiology being found. This result is consistent with multiple reports in the pediatric literature associating fever height with a prediction of infection in FN and ICU transfer.^25-27 Notably, similar literature is limited in adults with fever (who potentially have less robust immune responses) but emphasizes the notion that fever should not necessarily be considered a binary variable in our clinical decision-making. The importance of VS is also relevant in light of the increased use and evaluation of wearable VS monitoring devices in both nonmedical (eg, Fitbit) and medical settings. Wearable devices are positioning themselves in the field of oncology to identify toxicities from chemotherapy, assist in monitoring patients in the outpatient setting, and identify which patients are likely to get sicker after certain types of therapy, such as early detection of cytokine release syndrome after chimeric antigen receptor T-cell therapy.²⁸ Although the explosive growth of digital technology has many potential benefits, ongoing challenges include the management and interpretation of large volumes of physiologic data. Risk prediction models will ideally be integrated into this massive information output to help clinicians focus on the most essential data when making clinical assessments.

In addition to VS, we expected that baseline comorbidities would be significant predictors of FN outcomes, not only because they are included in all the other FN risk models but also because they are available at presentation, a critical time for decision-making. However, no baseline comorbidity that we recorded was associated with LR admissions in univariate or multivariate analysis, and further, the CCI, a validated index combining comorbidities, did not predict LR vs HR admission (AUC = 0.62) or infection identified (AUC = 0.51) well in this population. The reason comorbidities were less predictive in this AML population is not known, but it could be because we limited our population to those receiving intensive therapy who, in being eligible for intensive therapy, is an inherently healthier population with fewer comorbidities. For example, only 3% of the cohort had chronic obstructive pulmonary disease and 5% had a history of congestive heart failure. Therefore, comorbidities may be more predictive in the more heterogenous group of patients who receive less intensive chemotherapy.

Finally, we expected age to be a significant factor for the identification of LR admissions based on many of the currently available FN and ICU outcome models in which it is included. Interestingly, although age was significant in both the univariate and multivariate models for LR admission, it was in the reverse direction of the established models, with older age associated with LR admissions. Age was not associated with an infectious etiology identified as a cause for the FN. In analyzing differences between patients aged ≤65 years and those aged >65 years, we found that older patients had some HR features as expected (eg, higher PS and more cardiac comorbidities); none of these differences explain the association between older age and LR admissions. Furthermore, there were no differences between these cohorts in other comorbidities, other baseline laboratories, or physiologic variables, or whether infectious etiology was identified that could explain the association we found. It is possible that our cohort was too small to identify the causative difference, or that the older patients in this study likely represent a “healthier” subset of the older population with fewer comorbidities compared with older patients in a general oncology population (given that they were deemed fit enough to receive intensive chemotherapy), allowing for less complicated admissions. Finally, unrecorded factors that contributed to our outcome definition (eg, older patients may be less likely to want ICU transfer if they present with severe illness) may have led to this finding, which should be further explored using a larger data set.

Importantly, a FN risk prediction model will be most useful in distinguishing which patients may safely be treated in the outpatient setting if it uses presenting variables alone, rather than 24-hour variables, allowing for rapid decision-making in the clinic or ED. Our MV models incorporating the most abnormal 24-hour variables outperformed those using only presenting variables, unsurprisingly, as they allow more time and data to reflect the severity of illness. More work needs to be done on our AML-specific models to help improve the prediction at the time of the patient’s initial presentation with fever. One interesting approach to enhance this initial evaluation would be to incorporate existing and novel biomarkers specific to patients with hematologic malignancy undergoing intensive chemotherapy. Although existing critical illness biomarkers (eg, C-reactive protein) may perform more poorly in the patients who are immunosuppressed,²⁹ and newer biomarker candidates are not yet ready for clinical use, their integration into clinical risk models ultimately may improve the predictive ability of vital sign-based models and allow for earlier therapeutic decision-making in FN. Finally, in addition to the use of presenting variables, given the high risk of morbidity and mortality associated with FN, a predictive model of this nature is only valuable if the AUC is very high, likely near 1.0 in this case, given the potential high risk of predicting incorrectly. Even our best model (using 24-hour variables within a single data set) had an AUC of 0.82 for FN risk, which is not adequate given potential risks associated with undertreated FN. Thus, our data suggest that although some LR cases of FN may exist in solid malignancies, all FN in AML should still be considered high-risk, and the current practice to admit all patients with AML presenting with FN for rapid treatment and evaluation is still justified.

Several limitations of our study warrant consideration. First, as a retrospective cohort study, our analysis is subject to inherent biases associated with the retrospective data collection. Second, our study was conducted at a single center, and our institution has a relatively unique care model for managing many postinduction patients with AML in an outpatient setting,^14,30 which may limit the generalizability of our findings. Next, we only included patients who had received high-intensity treatment for AML; therefore, our results cannot be extrapolated to patients receiving lower-intensity therapies given potential differences in patient- and therapy-associated factors. Additionally, although 1 of our outcomes was whether an infectious etiology was identified, often the diagnosis of infection in AML is limited by the risks of the diagnostic procedure in our pancytopenic population (and thus the procedure is not done) and the limits of our diagnostic technology, leading some patients with true infection to be misidentified. Finally, our LR cohort was relatively small, which may have limited our ability to identify predictive factors. This may have been due to how we defined a LR admission, which was centered on the idea that a shorter admission (without death or ICU care need) is associated with less severe illness. Typically, when patients with AML present with FN at our center, we monitor them for 48 hours to rule out bacteremia and ensure that fevers resolve and then discharge them the following day, if this is the case. Some patients without complicated FN may stay longer because of logistics but not illness severity. We thus also re-evaluated our data using a 96-hour rather than a 72-hour definition of LR outcomes, which while increasing the number of LR admissions to 35%; it did not change the general results or discriminatory ability of our models. As a next step, we will evaluate alternative outcome measures to identify and define a LR subset of patients.

In conclusion, our study provides novel insights into how best to predict which patients with AML presenting with FN are less likely to develop a severe illness. By developing a MV model, we identified key predictors of LR admissions and demonstrated the limitations of the existing FN risk models in this population. Our findings underscore the need for tailored risk prediction tools to guide clinical decision-making and optimize the management of FN in patients with AML. Ideally, such a tool could be used to identify selected patients with AML in whom close outpatient management of FN with IV antibiotics and advanced VS monitoring tools would be safe, thus alleviating the burden of hospitalization and improving patient quality of life. However, such a tool would require almost 100% predictive accuracy given the high risk of incorrect predictions. Based on our results, much work still needs to be done to develop such a model for AML, and the current approach to hospitalization should remain the standard of care. Future research will aim to refine our LR outcome measures and further develop and validate our model in multicenter cohorts, including those treated with lower-intensity therapies, to enhance its generalizability, assess its performance in diverse clinical settings, and assess how to best integrate it into clinical practice.

R.B.W. acknowledges support from the José Carreras/E. Donnall Thomas Endowed Chair for Cancer Research.

Contribution: K.V.P., R.B.W., and A.B.H. designed and carried out research, analyzed data, and wrote the manuscript; M.O. analyzed data and critically edited the manuscript; Z.A. designed the research and collected data; K.R. and C. S. collected data and critically edited the manuscript; and M.-E.M.P., P.C.H., and J.S.A. designed the research and critically edited the manuscript.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Anna B. Halpern, Division of Hematology and Oncology, Department of Medicine, University of Washington, 825 Eastlake Ave E, Box LG-700, Seattle, WA 98109; email: halpern2@uw.edu.