Comments
Abstract
Rationale: Despite etiologic and severity heterogeneity in neutropenic sepsis, management is often uniform. Understanding host response clinical subphenotypes might inform treatment strategies for neutropenic sepsis.
Objectives: In this retrospective two-hospital study, we analyzed whether temperature trajectory modeling could identify distinct, clinically relevant subphenotypes among oncology patients with neutropenia and suspected infection.
Methods: Among adult oncologic admissions with neutropenia and blood cultures within 24 hours, a previously validated model classified patients’ initial 72-hour temperature trajectories into one of four subphenotypes. We analyzed subphenotypes’ independent relationships with hospital mortality and bloodstream infection using multivariable models.
Measurements and Main Results: Patients (primary cohort n = 1,145, validation cohort n = 6,564) fit into one of four temperature subphenotypes. “Hyperthermic slow resolvers” (pooled n = 1,140 [14.8%], mortality n = 104 [9.1%]) and “hypothermic” encounters (n = 1,612 [20.9%], mortality n = 138 [8.6%]) had higher mortality than “hyperthermic fast resolvers” (n = 1,314 [17.0%], mortality n = 47 [3.6%]) and “normothermic” (n = 3,643 [47.3%], mortality n = 196 [5.4%]) encounters (P < 0.001). Bloodstream infections were more common among hyperthermic slow resolvers (n = 248 [21.8%]) and hyperthermic fast resolvers (n = 240 [18.3%]) than among hypothermic (n = 188 [11.7%]) or normothermic (n = 418 [11.5%]) encounters (P < 0.001). Adjusted for confounders, hyperthermic slow resolvers had increased adjusted odds for mortality (primary cohort odds ratio, 1.91 [P = 0.03]; validation cohort odds ratio, 2.19 [P < 0.001]) and bloodstream infection (primary odds ratio, 1.54 [P = 0.04]; validation cohort odds ratio, 2.15 [P < 0.001]).
Conclusions: Temperature trajectory subphenotypes were independently associated with important outcomes among hospitalized patients with neutropenia in two independent cohorts.
The heterogeneity of sepsis, a syndromic illness identified by clinical parameters, offers the opportunity to improve outcomes through tailored management if clinically meaningful subphenotypes can be reproducibly identified. In the general sepsis population, longitudinal temperature trajectory models have reliably identified distinct subphenotypes with differential immune profiles and clinical outcomes. Such an approach has not been explored in patients with neutropenia but may be relevant for earlier recognition and treatment given that febrile neutropenia is neither sensitive nor specific for infection.
We validated four sepsis subphenotypes in two cohorts of hospitalized oncology patients with neutropenia and suspected infection on presentation. These groups had different outcomes, including hospital mortality, critical care use, bloodstream infections, and hospital lengths of stay. Consistently, outcomes were worse among “hyperthermic slow resolvers” and “hypothermic” encounters. We also trained and externally validated a machine learning model aimed at identifying hyperthermic slow resolvers earlier in their course, thus taking a key step toward clinical evaluation of this approach.
Neutropenic sepsis is associated with high morbidity and mortality in patients with malignancy (1). In patients with neutropenia, mucosal translocation of colonizing microbes may establish an infection for which typical early clinical signs are muted (2); fever may be the only manifestation of bacteremia (3). Given these patients’ immunocompromised state, it is critical to recognize neutropenic sepsis early to prevent clinical deterioration.
Although the risks of neutropenic febrile syndromes are well known, current triage and management approaches are suboptimal. The Infectious Disease Society of America guidelines define febrile neutropenia on the basis of a single hour’s temperature measurement and recommend blood cultures and antibiotics for patients with neutropenia meeting this definition (4). However, fever may be insensitive for systemic infection, as even temperature fluctuations below the fever threshold are associated with increased mortality (4–6). Oncology patients frequently receive therapies that may attenuate (e.g., glucocorticoids) (7) or cause fevers (e.g., blood products), limiting both the sensitivity and specificity of a single temperature threshold to predict outcomes. Thus, afebrile neutropenic sepsis may have delayed recognition and treatment (8). In contrast, unnecessary antimicrobial exposure in response to noninfectious fevers risks multidrug resistant organisms and medication-related complications without an associated benefit (9–11).
Better methods are needed to risk stratify patients with neutropenia with potential infection. In the general sepsis population, temperature trajectory models have reproducibly identified distinct subphenotypes with differential immune profiles and clinical outcomes (12, 13). Such an approach has not been explored in patients with neutropenia but may help identify patients at high risk for infection-related decompensation who do not meet the current definition of febrile neutropenia. We thus hypothesized that temperature trajectory modeling would categorize patients with neutropenia and suspected infection into subphenotypes with distinct clinical characteristics and outcomes. Some of these results have been previously reported in the form of an abstract (14).
This was a retrospective cohort study, reported in accordance with Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (see the online supplement) (15). This analysis was approved by Institutional Review Boards at Washington University School of Medicine and University of Chicago Medicine. Additional methodology details can be found in the online supplement.
In the primary cohort, we analyzed all hospital admissions to the oncology pavilion at Barnes-Jewish Hospital (BJH), a 1,300-bed tertiary-care hospital, between January 1, 2014, and June 30, 2017 (16). In an external validation cohort, we analyzed all oncology admissions at University of Chicago Medicine, an 800-bed tertiary-care hospital, between January 1, 2014, and December 31, 2019. In both cohorts, inclusion criteria were adult patients (age ≤ 18 yr) with cancer or stem cell transplantation who had both neutropenia (absolute neutrophil count [ANC] ⩽ 1,500 cells/mm3) and suspected infection (blood cultures collected) within 24 hours of admission (17, 18). We excluded patients with hospital length of stay (LOS) <24 hours, as they lacked sufficient temperature measurements for trajectory modeling.
We collected demographic data from the electronic health record. We determined Elixhauser comorbidities and the van Walraven index using billing data (19). As per prior methods, we used billing data to categorize patients into one of four malignancy categories: leukemia, lymphoma, multiple myeloma, or solid tumors (16).
Because commonly used severity-of-illness surrogates require information unavailable in this data set (e.g., mental status determinations for Sequential Organ Failure Assessment [SOFA]), would have confounded temperature trajectories (e.g., temperature for Modified Early Warning Score), or were likely to be significantly correlated with temperature extremes (e.g., heart rate in shock index), we used minimum systolic blood pressure (SBP) within the first 24 hours as a surrogate for severity of illness (20, 21). We extracted all temperature measurements from the encounter’s first 72 hours, together with daily minimum ANC measurements. Finally, we extracted medication orders for nonprophylactic antimicrobials (see Table E1 in the online supplement), granulocyte colony-stimulating factor (G-CSF), corticosteroids, and acetaminophen from the BJH cohort. To these, we added blood culture results and antimicrobial susceptibilities.
The primary outcome was hospital mortality. Secondary outcomes were ICU admission, hospital LOS, and bloodstream infections (BSIs). Within the primary cohort, additional secondary outcomes were receipt of vasoactive medications, receipt of renal replacement therapy, and time from initial blood cultures to antimicrobial initiation.
Using temperature data from each admission’s first 72 hours, a previously validated group-based trajectory modeling algorithm categorized patients into one of four subphenotypes: “hyperthermic slow resolvers” (HSRs), “hyperthermic fast resolvers” (HFRs), “normothermic,” and “hypothermic” (12). Here, hyperthermic, normothermic, and hypothermic are not defined by particular thresholds but rather describe the general relative temperature patterns observed across the subphenotypes.
We used chi-square, Fisher exact, and Kruskal-Wallis tests as appropriate for unadjusted comparisons.
To evaluate the independent relationships between subphenotype membership and binary outcomes, at each hospital we fit multivariable logistic regression models for mortality, ICU admission, and BSI. Analogously, to estimate the independent relationship between subphenotype membership and hospital LOS, we fit an extended multivariable Fine-Gray survival model, which treated death as a competing risk for discharge alive (22, 23). In this model, the subdistribution hazard ratio (SHR) conveys the relative probability of discharge alive at any time point (i.e., SHR < 1 indicates a lower adjusted probability of discharge, suggesting a longer adjusted LOS).
We performed several sensitivity analyses in the primary cohort to ensure the robustness of our findings. First, we repeated multivariable models within the narrower cohort of patients with severe neutropenia (ANC ⩽ 500 cells/mm3 within 24 h). Second, we refit models within the subset of patients who received nonprophylactic antimicrobials within 24 hours of admission. Third, we replicated these models using the 24-hour maximum value of a modified SOFA score as a more comprehensive marker of illness severity than SBP (24). Finally, on the basis of preliminary findings, we conducted a post hoc analysis in which adjusted models also included corticosteroids within the first 72 hours of hospitalization as a potential confounder (25).
To determine the predictive value added through temperature trajectory subphenotyping, we constructed a multivariable logistic regression model for hospital mortality in which predictors were the available clinically relevant information through hour 72. Using the likelihood ratio test, we compared this model to an otherwise identical model that also included the subphenotype as a predictor.
Finally, to increase the potential clinical utility of the primary model (26), we used data from the initial 24 hours of hospitalization to train and evaluate a parsimonious machine learning classifier model to predict ultimate membership in the HSR subgroup. In this model, prespecified variables were the patient’s age and gender, malignancy category, the minimum and maximum temperature, the minimum SBP, the maximum heart rate, the minimum ANC, and the total acetaminophen dose.
We trained the classifier model in the primary cohort using the XGBoost algorithm (27). We estimated classifier discrimination (area under the receiver operating characteristic curve [AUROC]) internally (primary) and externally (validation) using separate 2,000 bootstrap replications of each cohort (28). We used Shapley additive explanations to depict each variable’s contribution to classifying HSR subphenotype membership (29).
We performed all analyses using R version 4.1.2 with the tidyverse, cmprsk, and XGBoost packages (https://www.r-project.org/) (27, 30, 31). We considered two-sided P values <0.05 to indicate statistical significance.
In the primary cohort, 1,172 encounters had both neutropenia and suspicion of infection during the first 24 hours. Of these, 27 had hospital LOS <24 hours and were excluded from analysis, yielding 1,145 encounters. After grouping via the temperature trajectory algorithm, a plurality was categorized as normothermic (n = 427 [37%]), followed by HSRs (n = 275 [24%]), HFRs (n = 245 [21%]), and hypothermic (n = 198 [17%]) (Figure 1). In the validation cohort, 6,564 encounters met inclusion criteria after the exclusion of 158 with LOS <24 hours. Temperature trajectory modeling yielded a similar distribution of subphenotype membership but with relatively fewer HSR encounters and more normothermic encounters (normothermic, n = 3,216 [49%]; HSRs, n = 865 [13%]; HFRs, n = 1,069 [16%]; hypothermic, n = 1,414 [22%]).
Figure 1. Temperature subphenotypes: mean temperature over time. A group-based trajectory model produced four temperature trajectory subphenotypes within the neutropenic cohort. Smoothed curves represent mean hourly temperatures within each subphenotype, while the surrounding lightly shaded curves represent the 95% confidence intervals around these means. HFR = hyperthermic fast resolver; HSR = hyperthermic slow resolver.
[More] [Minimize]Patient characteristics for each cohort are listed in Table 1. In general, the cohorts were similar except that in the primary cohort, more patients identified as White, and the overall comorbidity burden was lower. In the primary cohort, the most common underlying malignancy was leukemia (n = 593 [52%]), whereas in the validation cohort the most common malignancy was solid tumor (n = 3,701 [56%]). At both sites, the prevalence rates of malignancy differed significantly across subphenotypes (BJH, P = 0.009; University of Chicago Medicine, P < 0.001). In the primary cohort, hypothermic (n = 118 [60%]) encounters had the highest proportion of patients with leukemia, while in the validation cohort, hypothermic encounters had the lowest proportion of patients with leukemia (n = 406 [29%]).
| Primary Cohort (n = 1,145) | Validation Cohort (n = 6,564) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| HSRs (n = 275) | HFRs (n = 245) | NT (n = 427) | HT (n = 198) | P Value | HSRs (n = 865) | HFRs (n = 1,069) | NT (n = 3,216) | HT (n = 1,414) | P Value | |
| Age, yr, median (IQR) | 58 (48–67) | 60 (50–67) | 62 (52–69) | 61 (52–70) | 0.010 | 59 (45–67) | 61 (48–69) | 63 (52–71) | 31 (53–72) | <0.001 |
| Female, n (%) | 116 (42.2) | 109 (44.4) | 196 (45.9) | 83 (41.9) | 0.71 | 355 (41.0) | 489 (45.7) | 1,544 (48.0) | 693 (53.5) | <0.001 |
| Race, n (%) | 0.26 | <0.001 | ||||||||
| White | 215 (78.2) | 201 (82.0) | 363 (85.0) | 170 (85.9) | 465 (53.8) | 547 (51.2) | 1,462 (45.5) | 632 (44.7) | ||
| Black | 40 (14.5) | 31 (12.7) | 45 (10.5) | 21 (10.6) | 288 (33.3) | 417 (39.0) | 1,504 (46.8) | 693 (49.0) | ||
| Other | 20 (7.3) | 13 (5.3) | 19 (4.4) | 7 (3.5) | 112 (12.9) | 105 (9.8) | 250 (7.8) | 89 (6.3) | ||
| Comorbidities | ||||||||||
| CHF, n (%) | 43 (15.6) | 40 (16.3) | 85 (19.9) | 49 (24.7) | 0.054 | 121 (14.0) | 194 (18.1) | 685 (21.3) | 375 (25.6) | <0.001 |
| Hypertension, n (%) | 189 (69.5) | 170 (69.4) | 331 (55.2) | 165 (83.3) | 0.59 | 586 (67.7) | 858 (80.3) | 2,812 (87.4) | 1,288 (91.1) | <0.001 |
| Chronic lung disease, n (%) | 39 (14.2) | 52 (21.2) | 108 (25.3) | 50 (23.5) | 0.003 | 214 (24.7) | 273 (25.5) | 1,004 (31.2) | 466 (33.0) | <0.001 |
| Diabetes, n (%) | 91 (33.1) | 92 (37.6) | 156 (36.5) | 75 (37.9) | 0.66 | 60 (21.7) | 53 (21.3) | 102 (23.0) | 50 (24.6) | <0.001 |
| Hypothyroid, n (%) | 53 (19.3) | 44 (18.0) | 65 (15.2) | 40 (20.2) | 0.37 | 131 (15.1) | 194 (18.1) | 599 (18.6) | 273 (19.3) | 0.07 |
| Renal failure, n (%) | 48 (17.5) | 46 (18.8) | 81 (19.0) | 57 (28.8) | 0.012 | 153 (17.7) | 238 (22.3) | 780 (24.3) | 360 (25.5) | <0.001 |
| Liver disease, n (%) | 36 (13.1) | 26 (11.0) | 53 (12.4) | 27 (13.6) | 0.85 | 140 (16.2) | 216 (20.2) | 602 (18.7) | 320 (22.6) | 0.001 |
| vW score, median (IQR) | 15 (8.5–23) | 17 (9–25) | 17 (9–25) | 20 (12–27) | 0.023 | 20 (11–28) | 23 (14–31) | 24 (15–33) | 26 (16–34) | <0.001 |
| vW score, without cancer, median (IQR) | 11 (5–18) | 11 (5–17) | 11 (5–18) | 14 (7–22) | 0.01 | 13 (5–20) | 13 (6–20) | 14 (7–21) | 14 (7–22) | <0.001 |
| Underlying cancer, n (%) | 0.009 | <0.001 | ||||||||
| Leukemia | 147 (53.5) | 110 (44.9) | 218 (51.0) | 118 (59.5) | 422 (48.8) | 425 (39.8) | 1,118 (34.8) | 406 (28.7) | ||
| Lymphoma | 59 (21.5) | 61 (24.9) | 73 (17.1) | 34 (17.2) | 43 (5.0) | 48 (4.5) | 97 (3.0) | 73 (5.2) | ||
| Myeloma | 3 (1.1) | 5 (2.0) | 11 (2.6) | 10 (5.1) | 46 (5.3) | 53 (5.0) | 84 (2.6) | 48 (3.4) | ||
| Solid tumor | 61 (22.2) | 66 (26.9) | 118 (27.6) | 35 (17.7) | 354 (40.9) | 543 (50.8) | 1,917 (59.6) | 887 (62.7) | ||
| Unknown | 5 (1.8) | 3 (1.2) | 7 (1.6) | 1 (0.5) | 422 (48.8) | 425 (39.8) | 1,118 (34.8) | 406 (28.7) | ||
| Physiology during the initial 24 h | ||||||||||
| Minimum SBP, mm Hg, median (IQR) | 103 (92–113) | 102 (92–112) | 103 (95–112) | 105 (95–118) | 0.28 | 99 (91–110) | 99 (90–110) | 102 (92–114) | 101 (91–113) | <0.001 |
| Minimum ANC, ×109/L, median (IQR) | 0.1 (0.0–0.5) | 0.2 (0.0–0.7) | 0.3 (0.1–0.8) | 0.5 (0.2–1.0) | <0.001 | 0.3 (0.1–0.8) | 0.4 (0.1–0.9) | 0.6 (0.3–1.0) | 0.6 (0.3–1.0) | <0.001 |
Temperatures varied as expected across subphenotypes during the first 72 hours of admission (Figure 1; see Table E2). ANC on admission varied across subphenotypes as well (Table 1; see Figure E1), with the HSR subphenotype having the lowest and hypothermic encounters having the highest values in both cohorts (P < 0.001). Duration of neutropenia was longer by approximately 2 days for the HSR subphenotype (median, 7 [interquartile range (IQR), 4–13] d) than in the HFR (median, 4 [IQR, 3–7] d), normothermic (median, 4 [IQR, 2–9] d), and hypothermic (median, 4.5 [IQR, 3–9] d) groups (P < 0.001).
Table E3 delineates medication orders by subphenotype in the primary cohort. Acetaminophen orders were higher in the HSR group than for the other subphenotypes throughout this period (P < 0.001; see Figure E2). Corticosteroid orders differed significantly among temperature trajectories (P < 0.001), with the highest rates occurring among hypothermic (n = 99 [50%]) and HFR (n = 86 [35%]) encounters. Orders for G-CSF were similar across subphenotypes, except for HSRs, who were more likely to receive G-CSF at 48–72 hours.
Most patients (n = 1,069 [93%]) had orders for antimicrobials within the first 24 hours of hospitalization; encounters in the hyperthermic subphenotypes (HSRs, n = 268 [97.4%]; HFRs, n = 235 [95.9%]) were more likely than normothermic (n = 391 [91.6%]) or hypothermic (n = 175 [88.4%]) encounters to have received antibiotics in this window (P < 0.001). Among patients who received antimicrobials, the median time between blood cultures and the first nonprophylactic antimicrobial order (with negative values indicating that antimicrobials were ordered before cultures were collected) ranged from −0.2 hours (IQR, −0.9 to 2.5 h) in HFRs to −0.5 hours (IQR, −1.3 to 0.9 h) among patients with normothermia (P = 0.006).
In the primary cohort, hospital death occurred in 78 (6.8%) encounters (Table 2). Mortality differed significantly among subphenotypes (P = 0.004), with 26 (9.5%) in-hospital deaths for the HSR subphenotype, followed by patients with hypothermia (n = 20 [10.1%]) and those with normothermia (n = 24 [5.6%]). Similar patterns occurred for ICU admissions, vasoactive medications, and renal replacement therapy (see Table E4). LOS (Figure 2) was significantly higher in the HSR group (median, 8 [IQR, 5–15] d) than the HFR (median, 5 [IQR, 3–8] d), normothermic (median, 5 [IQR, 3–11] d), and hypothermic (median, 6 [IQR, 3–12] d) subphenotypes (P < 0.001). These findings were similar in the validation cohort.
| Primary Cohort (n = 1,145) | Validation Cohort (n = 6,564) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| HSRs (n = 275) | HFRs (n = 245) | NT (n = 427) | HT (n = 198) | P Value | HSRs (n = 865) | HFRs (n = 1,069) | NT (n = 3,216) | HT (n = 1,414) | P Value | |
| In-hospital death, n (%) | 26 (9.5) | 8 (3.3) | 24 (5.6) | 20 (10.1) | 0.007 | 78 (9.0) | 39 (3.6) | 172 (5.3) | 118 (8.3) | <0.001 |
| ICU admission, n (%) | 59 (21.5) | 39 (15.9) | 51 (11.9) | 29 (14.6) | 0.009 | 226 (26) | 172 (16) | 675 (21) | 287 (20) | <0.001 |
| LOS, d, median (IQR) | 8 (5–15) | 5 (3–8) | 5 (3–11) | 6 (3–12) | <0.001 | 7 (4–14) | 5 (3–8) | 5 (3–9) | 5 (3–9) | <0.001 |
| BSI, n (%) | 55 (19.9) | 52 (20.9) | 56 (12.6) | 23 (11.3) | 0.002 | 193 (22) | 188 (18) | 362 (11) | 165 (12) | <0.001 |
Figure 2. The cumulative incidences of death and hospital discharge varied by temperature trajectory subphenotype. The x-axis represents time in the hospital, while the y-axis represents the proportion of encounters having met one of two mutually exclusive endpoints: death (dotted lines) or discharge alive (solid lines). Each subphenotype is indicated by a different color. HFR = hyperthermic fast resolver; HSR = hyperthermic slow resolver.
[More] [Minimize]BSI rates varied among subphenotypes (P = 0.002), with 55 (20.0%) HSRs and 52 (21.2%) HFRs having one or more positive blood cultures in the primary cohort. Normothermic and hypothermic encounters had 56 (12.6%) and 23 (11.6%) BSIs, respectively. Most BSIs were bacteremic (see Table E5). Rates of candidemia (P = 0.997), gram-negative (P = 0.096), methicillin-resistant Staphylococcus aureus (P = 0.372), and vancomycin-resistant Enterococcus (P = 0.697) did not differ among subphenotypes, but extended-spectrum β lactamase–producing gram-negative bacteremia was statistically more common among HSRs (n = 14 vs. n = 7 or 8 among other subphenotypes; P = 0.042). Viremia was uncommon (n = 48) and did not differ among subphenotypes. Notably, 70 (38%) BSIs occurred in patients whose temperatures never reached 38°C, including 1 for the HSR subphenotype, 13 for the HFR subphenotype, 35 in normothermic encounters, and 21 in hypothermic encounters.
Of patients with BSIs, the time between culture collection and orders for appropriate antimicrobials (i.e., agents to which laboratory testing demonstrated susceptibility) did not differ among subphenotypes (Figure 3; P = 0.60).
Figure 3. Among encounters with culture-proven bloodstream infection, time from culture to orders for appropriate antimicrobials did not differ by subphenotype. The y-axis represents time from culture collection to the first order for appropriate antimicrobial therapy, on the basis of individual culture susceptibilities. Each box/whisker plot indicates the median (middle horizontal line), interquartile range (IQR; lower and upper “hinges” outlining the box), and extreme values (“whiskers” cover 75% of the data). The overlying scatterplot indicates individual observations. Among HSRs, median time to appropriate treatment was 13 hours (IQR, 6–25 h), which did not statistically differ from that of HFRs (median, 10 [IQR, 5–17] h), normothermic encounters (median, 11 [IQR, 5–62] h), or hypothermic encounters (median, 13 [IQR, 6–25 h]) (P = 0.60). HFR = hyperthermic fast resolver; HSR = hyperthermic slow resolver.
[More] [Minimize]In both cohorts, after adjusting for prespecified confounders, the HSR subphenotype was associated with significantly higher in-hospital mortality compared with patients with normothermia (Table 3; primary: adjusted odds ratio [aOR], 1.86 [95% confidence interval (CI), 1.07–3.41; P = 0.04]; validation: aOR, 2.19 [95% CI, 1.63–2.94; P < 0.001]). There were significantly higher adjusted odds for BSIs among HSRs (primary: aOR, 1.54 [95% CI, 1.01–2.34; P = 0.04]; validation: aOR, 2.15 [95% CI, 1.76–2.62; P < 0.001]) and HFRs (primary: aOR, 1.85 [95% CI, 1.21–2.81; P = 0.004]; validation: aOR, 1.60 [95% CI, 1.32–1.94; P < 0.001]) subphenotypes compared with normothermic encounters. Finally, the SHR for the HSR subphenotype was consistent with longer adjusted LOS in both cohorts (primary: SHR, 0.67 [95% CI, 0.57–0.77; P < 0.001]; validation: SHR, 0.69 [95% CI, 0.64–0.74; P < 0.001]). In contrast, the HFR subphenotype (primary: SHR, 1.19 [95% CI, 1.01–2.41; P = 0.04]; validation: SHR, 1.12 [95% CI, 1.05–1.20; P = 0.001]) had a significantly shorter adjusted LOS.
| Primary Cohort (n = 1,145) | Validation Cohort (n = 6,564) | |||||
|---|---|---|---|---|---|---|
| HSRs (n = 275) | HFRs (n = 245) | HT (n = 198) | HSRs (n = 865) | HFRs (n = 1,069) | HT (n = 1,414) | |
| Dependent variable | Multivariable logistic regression | |||||
| In-hospital mortality, aOR (95% CI) | 1.91 (1.07–3.41), P = 0.03 | 0.58 (0.24–1.25), P = 0.18 | 1.70 (0.90–3.16), P = 0.09 | 2.19 (1.63–2.94), P < 0.001 | 0.73 (0.50–1.04), P = 0.09 | 1.57 (1.22–2.01), P < 0.001 |
| ICU admission, aOR (95% CI) | 2.01 (1.30–3.13), P = 0.002 | 1.38 (0.86–2.20), P = 0.18 | 1.25 (0.74–2.10), P = 0.39 | 2.85 (1.73–4.69), P < 0.001 | 0.72 (0.60–0.89), P < 0.001 | 0.91 (0.77–1.07), P = 0.25 |
| Bloodstream infection, aOR (95% CI) | 1.54 (1.01–2.34), P = 0.04 | 1.85 (1.21–2.81), P = 0.004 | 0.88 (0.51–1.47), P = 0.63 | 2.15 (1.76–2.62), P < 0.001 | 1.60 (1.32–1.94), P < 0.001 | 1.02 (0.84–1.24), P = 0.82 |
| Multivariable Fine-Gray competing-risks model | ||||||
| Length of stay, SHR* (95% CI) | 0.67 (0.57–0.77), P < 0.001 | 1.19 (1.01–1.41), P = 0.04 | 0.87 (0.72–1.06), P = 0.16 | 0.69 (0.64–0.74), P < 0.001 | 1.12 (1.05–1.20), P < 0.001 | 0.92 (0.86–0.98), P = 0.008 |
The likelihood ratio test indicated that on the basis of multivariable logistic regression modeling, subphenotypic identification carried additional predictive value (P = 0.014) beyond that of demographics, comorbidities, neutropenia severity and duration, SOFA score, time to antimicrobial orders, and BSI.
Each of the sensitivity analyses yielded somewhat similar results to the primary analysis (see Table E6). Among the 750 encounters with severe neutropenia on admission, aOR point estimates were similar to those estimated from the entire primary cohort but were not statistically significant. Of those encounters with antimicrobials ordered during the first 24 hours (n = 1,069), both the HSR and HFR subphenotypes were associated with higher adjusted odds of BSI (HSR: aOR, 1.57 [95% CI, 1.02–2.43; P = 0.04]; HFR: aOR, 1.89 [95% CI, 1.22–2.94; P = 0.004]). Nearly identical BSI findings occurred in the analysis of the entire cohort adjusted for SOFA score instead of SBP (HSR: aOR, 1.55 [95% CI, 1.01–2.39; P = 0.045]; HFR: aOR, 1.87 [95% CI, 1.21–2.89; P = 0.005]). Finally, adjusting for corticosteroids in addition to the prespecified confounders produced results similar to those of the primary analysis; most notably, HSR encounters were again associated with increased adjusted odds for mortality (aOR, 2.02 [95% CI, 1.10–3.72; P = 0.02]).
The classifier model to predict HSR subphenotype membership using data from only the first 24 hours in hospitalization had nearly identical discrimination on internal (AUROC, 0.90 [95% CI, 0.88–0.93]) and external (AUROC, 0.90 [95% CI, 0.89–0.91]) validation. The Shapley values calculated from this model (Figure 4) indicated that first-day temperature measurements, cumulative acetaminophen dose, heart rate, and ANC were strongly contributory toward classifier accuracy.
Figure 4. Shapley additive explanation (SHAP) values indicate the relative contribution of each variable toward predicting hyperthermic slow resolver (HSR) membership in the classifier model. The x-axis represents the SHAP value (i.e., impact on model prediction), while the y-axis displays model variables, ranked from most predictive (bottom) to least predictive (top). Color coding indicates the degree to which higher or lower variable values lead to higher or lower predicted probabilities of HSR membership. ANC = absolute neutrophil count; HR = heart rate; SBP = systolic blood pressure.
[More] [Minimize]This study represents, to our knowledge, the first evaluation of a longitudinal modeling approach to evaluate the presence, characteristics, and outcomes of potential clinical subphenotypes among patients admitted to the hospital with neutropenia. These findings may have important implications for the triage and care of patients with neutropenia in the hospital.
Our study’s primary finding was that previously demonstrated temperature trajectory subphenotypes exist among patients with neutropenia and are independently associated with clinically important outcomes, including in-hospital mortality, BSIs, ICU use, and length of hospitalization. As has been demonstrated in several nonneutropenic cohorts, the hyperthermic slow-resolving subphenotype manifested the most severe outcomes (11, 12, 25, 32). These findings were consistent across two cohorts with different patient demographics, comorbidity burdens, and underlying malignancy prevalence rates. Importantly, these outcomes occurred despite the fact that HSRs were those most likely to have received antimicrobials within 24 hours, and that among those with BSIs, HSR encounters were no more likely than the other subphenotypes to involve significantly delayed time to appropriate antimicrobial treatments.
Our findings have translational value beyond their immediate clinical implications. For example, we found lower risk of in-hospital mortality and shorter hospital LOS in HFRs with severe neutropenia despite high rates of BSI, which is consistent with a prior temperature trajectory analysis of patients without neutropenia (12). The extent to which the underlying pathogen may contribute to these differences is unclear; HFRs generally had similar microbiologic patterns to HSRs in the primary cohort. This intriguing difference may suggest that HFRs had an inherently more adaptive response to infection despite initial neutropenia and temperatures similar to HSRs. Comparing measurements of inflammatory markers in similar patients may provide a better understanding of the divergent outcomes between the two subphenotypes.
Another possible mechanism for these findings may involve differences in immune composition over time, including prolonged exposure to cytokines, which have been shown to vary between subphenotypes (13). In addition, or alternatively, changes in HLA-DR expression over time (and therefore monocyte expression) have also been shown to differ between febrile and afebrile septic patients and to correlate with mortality (33). Finally, in a prior sepsis subphenotyping study, the hypothermic subphenotype had the lowest concentrations of G-CSF (13), which raises the hypothesis that neutrophil function could underlie a portion of this group’s biology (i.e., as an aspect of sepsis-induced immunosuppression) (34). By conditioning on neutropenia, our analysis has increased the focus on patients with hyperthermia and those with normothermia without typical healthy immune responses. It remains to be investigated whether differences in other immune cells’ function or clinical treatments might alter these clinical trajectories.
Relatedly, we quantified orders for corticosteroids, acetaminophen, and G-CSF, which represent potential modifiers of both temperature and neutropenia trends. Corticosteroid orders during the initial 72 hours were most common for HFR and hypothermic encounters, which opens the possibility that steroid administration may influence subphenotype membership, as suggested by prior work in patients with coronavirus disease (COVID-19) (25). Acetaminophen was ordered more frequently for patients with the HSR subphenotype than for other subphenotypes, suggesting that fevers occurred despite acetaminophen use in this group rather than because of underuse. Differential G-CSF ordering rates were modest and unlikely to affect 72-hour temperature trajectories, given that the median time to G-CSF response is 5–7 days (35).
A final important finding of our study is that many patients with neutropenic BSIs never experienced fevers. On the basis of the maximum recorded temperatures for patient encounters, current guidelines for neutropenic fever would have failed to identify 38% of BSIs in this cohort (4). Patients with normothermia and those with hypothermia also had higher rates of BSIs than commonly reported incidences in the literature (36). The implication of these findings is that overreliance on the current definition of neutropenic fever would have poor accuracy in identifying patients with BSIs (37). Alternative methods are needed to identify patients with neutropenia at risk for sepsis, to avoid morbidity and mortality associated with untreated or delayed treatment for BSIs (22).
One key goal of this subphenotyping approach is to provide low-cost prognostic (e.g., risk for mortality, LOS, BSI, likelihood of neutrophil recovery) and potentially theranostic information through a universally available measurement. For instance, transcriptomics-based endotyping (although neither universally available nor low cost at present) has identified evidence of heterogeneity of corticosteroid treatment effect in sepsis (38), and vital signs–based subphenotyping has recently identified differential responses to balanced crystalloids in sepsis (39). However, because temperature trajectory models generally require several days’ measurements for optimal accuracy (12, 25), their clinical value may be limited for sepsis recognition and initial management (26). To address this issue, we trained and externally validated a machine learning model able to identify HSRs, with high accuracy, 48 hours earlier than the temperature trajectory model itself. Such a model might facilitate timelier recognition, triage, and tailoring of sepsis treatments among patients with neutropenia and cancer, even in the absence of fever.
Key strengths of this study include its novel methodology within this patient population and its use of routinely collected noninvasive data over time to both risk stratify patients and generate new hypotheses regarding their differential clinical courses. Unlike other risk indices (e.g., the Multinational Association for Supportive Care in Cancer score) (40), which risk stratify patients only after the application of a binary classification of neutropenic fever, temperature trajectory modeling can evaluate all temperature measurements, regardless of any particular threshold. Adding rigor to this novel approach was the study’s prespecification of important confounders, use of competing risk analysis to account for LOS variation in patients who did versus those who did not die in the hospital. Measurement of important medications that could influence subphenotype membership, and consideration of multiple important clinical outcomes. Finally, the study’s conduct at two National Cancer Institute–designated comprehensive cancer centers allowed the inclusion of a broad and diverse set of patients, underlying malignancies, and spectra of treatment approaches, which may enhance generalizability.
Notwithstanding its strengths, this study has limitations. First, despite maximizing analytic rigor by specifying potential confounders a priori, the retrospective nature of this analysis leaves open the possibility of unmeasured confounding. For instance, because we were unable to determine the modality of temperature measurements (e.g., oral vs. tympanic), it is possible that differential measurements on the basis of clinical concern could have biased characterization of patients into particular subphenotypes. Second, the study’s lack of biological data (e.g., cytokine measurements) limits the inferences to be drawn regarding potential effect mechanisms. Finally, secondary use of electronic health record data prevents consideration of potentially important information such as the reason for neutropenia (e.g., planned stem cell transplantation, chemotherapy, marrow infiltration), the presence or duration of prehospital symptoms (including fevers, which may represent an opportunity for novel technological approaches to infection surveillance in high-risk patients) (41), or clinicians’ expectations regarding the severity and duration of neutropenia (which could have influenced treatment decisions). Accordingly, we did not have access to information related to patient goals of care or treatment limitations. Similarly, we were unable to ascertain clinical suspicion for sepsis, for which automated definitions have unknown validity among patients with neutropenia. However, this limitation also provides the advantage of allowing evaluation of all patients with neutropenia with any degree of suspicion of infection, including those who never manifested a fever.
We validated four previously described temperature trajectory subphenotypes among patients with neutropenia and cancer, extending prior research to a high-risk population. Moving toward the identification and study of treatment-responsive subgroups of neutropenic infection will be an important next step toward improved care and outcomes in neutropenia.
| 1. | Kuderer NM, Dale DC, Crawford J, Cosler LE, Lyman GH. Mortality, morbidity, and cost associated with febrile neutropenia in adult cancer patients. Cancer 2006;106:2258–2266. |
| 2. | Mortaz E, Alipoor SD, Adcock IM, Mumby S, Koenderman L. Update on neutrophil function in severe inflammation. Front Immunol 2018;9:2171. |
| 3. | Sickles EA, Greene WH, Wiernik PH. Clinical presentation of infection in granulocytopenic patients. Arch Intern Med 1975;135:715–719. |
| 4. | Freifeld AG, Bow EJ, Sepkowitz KA, Boeckh MJ, Ito JI, Mullen CA, et al.; Infectious Diseases Society of America. Clinical practice guideline for the use of antimicrobial agents in neutropenic patients with cancer: 2010 update by the Infectious Diseases Society of America. Clin Infect Dis 2011;52:e56–e93. |
| 5. | Gulleen EA, Ameko MK, Ainsworth JE, Barnes LE, Moore CC. Predictive models of fever, ICU transfer, and mortality in hospitalized patients with neutropenia. Crit Care Explor 2020;2:e0289. |
| 6. | Egi M, Morita K. Fever in non-neurological critically ill patients: a systematic review of observational studies. J Crit Care 2012;27:428–433. |
| 7. | Uda H, Suga Y, Toriba E, Staub AY, Shimada T, Sai Y, et al. Multiday corticosteroids in cancer chemotherapy delay the diagnosis of and antimicrobial administration for febrile neutropenia: a double-center retrospective study. J Pharm Health Care Sci 2019;5:3. |
| 8. | Strojnik K, Mahkovic-Hergouth K, Novakovic BJ, Seruga B. Outcome of severe infections in afebrile neutropenic cancer patients. Radiol Oncol 2016;50:442–448. |
| 9. | Baluch A, Shewayish S. Neutropenic fever. In: Velez AP, Lamarche J, Greene JN, editors. Infections in neutropenic cancer patients. Cham, Switzerland: Springer; 2019. pp. 105–117. |
| 10. | Gudiol C, Tubau F, Calatayud L, Garcia-Vidal C, Cisnal M, Sánchez-Ortega I, et al. Bacteraemia due to multidrug-resistant gram-negative bacilli in cancer patients: risk factors, antibiotic therapy and outcomes. J Antimicrob Chemother 2011;66:657–663. |
| 11. | Bodro M, Gudiol C, Garcia-Vidal C, Tubau F, Contra A, Boix L, et al. Epidemiology, antibiotic therapy and outcomes of bacteremia caused by drug-resistant ESKAPE pathogens in cancer patients. Support Care Cancer 2014;22:603–610. |
| 12. | Bhavani SV, Carey KA, Gilbert ER, Afshar M, Verhoef PA, Churpek MM. Identifying novel sepsis subphenotypes using temperature trajectories. Am J Respir Crit Care Med 2019;200:327–335. |
| 13. | Bhavani SV, Wolfe KS, Hrusch CL, Greenberg JA, Krishack PA, Lin J, et al. Temperature trajectory subphenotypes correlate with immune responses in patients with sepsis. Crit Care Med 2020;48:1645–1653. |
| 14. | Benzoni N, Bhavani S, Fuller B, Lyons PG. Temperature trajectory subphenotypes and associated outcomes among neutropenic patients [abstract]. Am J Respir Crit Care Med 2021;203:A1266. |
| 15. | von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Ann Intern Med 2007;147:573–577. |
| 16. | Lyons PG, Klaus J, McEvoy CA, Westervelt P, Gage BF, Kollef MH. Factors associated with clinical deterioration among patients hospitalized on the wards at a tertiary cancer hospital. J Oncol Pract 2019;15:e652–e665. |
| 17. | Trotti A, Colevas AD, Setser A, Rusch V, Jaques D, Budach V, et al. CTCAE v3.0: development of a comprehensive grading system for the adverse effects of cancer treatment. Semin Radiat Oncol 2003;13:176–181. |
| 18. | Churpek MM, Dumanian J, Dussault N, Bhavani SV, Carey KA, Gilbert ER, et al. Determining the electronic signature of infection in electronic health record data. Crit Care Med 2021;49:e673–e682. |
| 19. | van Walraven C, Austin PC, Jennings A, Quan H, Forster AJ. A modification of the Elixhauser comorbidity measures into a point system for hospital death using administrative data. Med Care 2009;47:626–633. |
| 20. | Subbe CP, Kruger M, Rutherford P, Gemmel L. Validation of a modified early warning score in medical admissions. QJM 2001;94:521–526. |
| 21. | Althunayyan SM, Alsofayan YM, Khan AA. Shock index and modified shock index as triage screening tools for sepsis. J Infect Public Health 2019;12:822–826. |
| 22. | Ji J, Klaus J, Burnham JP, Michelson A, McEvoy CA, Kollef MH, et al. Bloodstream infections and delayed antibiotic coverage are associated with negative hospital outcomes in hematopoietic stem cell transplant recipients. Chest 2020;158:1385–1396. |
| 23. | Austin PC, Fine JP. Practical recommendations for reporting Fine-Gray model analyses for competing risk data. Stat Med 2017;36:4391–4400. |
| 24. | Cetintepe T, Cetintepe L, Solmaz S, Calık S, Ugur MC, Gediz F, et al. Determination of the relationship between mortality and SOFA, qSOFA, MASCC scores in febrile neutropenic patients monitored in the intensive care unit. Support Care Cancer 2021;29:4089–4094. |
| 25. | Bhavani SV, Verhoef PA, Maier CL, Robichaux C, Parker WF, Holder A, et al. Coronavirus disease 2019 temperature trajectories correlate with hyperinflammatory and hypercoagulable subphenotypes. Crit Care Med 2022;50:212–223. |
| 26. | Prescott HC. Identifying sepsis subtypes from routine clinical data. Am J Respir Crit Care Med 2019;200:272–273. |
| 27. | Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: Association for Computing Machinery; 2016. pp. 785–794. |
| 28. | Steyerberg EW. Validation in prediction research: the waste by data splitting. J Clin Epidemiol 2018;103:131–133. |
| 29. | Lundberg SM, Erion G, Chen H, DeGrave A, Prutkin JM, Nair B, et al. From local explanations to global understanding with explainable AI for trees. Nat Mach Intell 2020;2:56–67. |
| 30. | Wickham H. Tidyverse [accessed 2022 May 10]. Available from: https://www.tidyverse.org. |
| 31. | Gray B. cmprsk: subdistribution analysis of competing risks. Comprehensive R Archive Network; 2019 [accessed 2022 May 10]. Available from: https://cran.r-project.org/web/packages/cmprsk/index.html. |
| 32. | Bhavani SV, Huang ES, Verhoef PA, Churpek MM. Novel temperature trajectory subphenotypes in COVID-19. Chest 2020;158:2436–2439. |
| 33. | Drewry AM, Ablordeppey EA, Murray ET, Dalton CM, Fuller BM, Kollef MH, et al. Monocyte function and clinical outcomes in febrile and afebrile patients with severe sepsis. Shock 2018;50:381–387. |
| 34. | Hotchkiss RS, Monneret G, Payen D. Immunosuppression in sepsis: a novel understanding of the disorder and a new therapeutic approach. Lancet Infect Dis 2013;13:260–268. |
| 35. | Greenbaum AM, Link DC. Mechanisms of G-CSF-mediated hematopoietic stem and progenitor mobilization. Leukemia 2011;25:211–217. |
| 36. | Schelenz S, Nwaka D, Hunter PR. Longitudinal surveillance of bacteraemia in haematology and oncology patients at a UK cancer centre and the impact of ciprofloxacin use on antimicrobial resistance. J Antimicrob Chemother 2013;68:1431–1438. |
| 37. | Kameda K, Kimura S-I, Akahoshi Y, Nakano H, Harada N, Ugai T, et al. High incidence of afebrile bloodstream infection detected by surveillance blood culture in patients on corticosteroid therapy after allogeneic hematopoietic stem cell transplantation. Biol Blood Marrow Transplant 2016;22:371–377. |
| 38. | Antcliffe DB, Burnham KL, Al-Beidh F, Santhakumaran S, Brett SJ, Hinds CJ, et al. Transcriptomic signatures in sepsis and a differential response to steroids: from the VANISH randomized trial. Am J Respir Crit Care Med 2019;199:980–986. |
| 39. | Bhavani SV, Semler M, Qian ET, Verhoef PA, Robichaux C, Churpek MM, et al. Development and validation of novel sepsis subphenotypes using trajectories of vital signs. Intensive Care Med 2022;48:1582–1592. |
| 40. | Klastersky J, Paesmans M, Rubenstein EB, Boyer M, Elting L, Feld R, et al. The Multinational Association for Supportive Care in Cancer risk index: a multinational scoring system for identifying low-risk febrile neutropenic cancer patients. J Clin Oncol 2000;18:3038–3051. |
| 41. | Nessle CN, Flora C, Sandford E, Choi SW, Tewari M. High-frequency temperature monitoring at home using a wearable device: a case series of early fever detection and antibiotic administration for febrile neutropenia with bacteremia. Pediatr Blood Cancer 2022; 69:e29835. |
* These authors contributed equally to this work.
‡ Co–senior authors.
Supported by NIH/National Center for Advancing Translational Sciences grant UL1 TR002345, NIH/National Institute of General Medical Sciences grant K23GM144867 (S.V.B.), NIH/National Center for Advancing Translational Sciences grant KL2 TR002346 (P.G.L.), and Doris Duke Charitable Foundation grant 2015215 (P.G.L.).
Author Contributions: N.S.B., K.A.C., A.F.B., J.K., B.M.F., D.P.E., M.M.C., S.V.B., and P.G.L. each made substantial contributions to the conception or design of the work and/or the acquisition, analysis, and interpretation of data for the work; drafted the work and revised it critically for important intellectual content; provided final approval of the version to be published; and are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.
Data Availability: Because of patient privacy concerns, supporting data cannot be made openly available. Supporting data may be made available to collaborators under a data use agreement with Washington University School of Medicine by contacting the corresponding author. Analysis code is available from the corresponding author upon request.
This article has an online supplement, which is accessible from this issue’s table of contents at www.atsjournals.org.
Originally Published in Press as DOI: 10.1164/rccm.202205-0920OC on November 30, 2022
Author disclosures are available with the text of this article at www.atsjournals.org.
