Comments
Abstract
Rationale: Sepsis is a heterogeneous syndrome, and identifying clinically relevant subphenotypes is essential.
Objectives: To identify novel subphenotypes in hospitalized patients with infection using longitudinal temperature trajectories.
Methods: In the model development cohort, inpatient admissions meeting criteria for infection in the emergency department and receiving antibiotics within 24 hours of presentation were included. Temperature measurements within the first 72 hours were compared between survivors and nonsurvivors. Group-based trajectory modeling was performed to identify temperature trajectory groups, and patient characteristics and outcomes were compared between the groups. The model was then externally validated at a second hospital using the same inclusion criteria.
Measurements and Main Results: A total of 12,413 admissions were included in the development cohort, and 19,053 were included in the validation cohort. In the development cohort, four temperature trajectory groups were identified: “hyperthermic, slow resolvers” (n = 1,855; 14.9% of the cohort); “hyperthermic, fast resolvers” (n = 2,877; 23.2%); “normothermic” (n = 4,067; 32.8%); and “hypothermic” (n = 3,614; 29.1%). The hypothermic subjects were the oldest and had the most comorbidities, the lowest levels of inflammatory markers, and the highest in-hospital mortality rate (9.5%). The hyperthermic, slow resolvers were the youngest and had the fewest comorbidities, the highest levels of inflammatory markers, and a mortality rate of 5.1%. The hyperthermic, fast resolvers had the lowest mortality rate (2.9%). Similar trajectory groups, patient characteristics, and outcomes were found in the validation cohort.
Conclusions: We identified and validated four novel subphenotypes of patients with infection, with significant variability in inflammatory markers and outcomes.
Sepsis is a heterogeneous syndrome, and identification of clinically relevant subphenotypes could lead to more personalized management. Body temperature responses are an immediately apparent heterogeneity in patients with infection and sepsis. Research has shown that hypothermia in the setting of infection is associated with increased mortality, whereas fever is associated with decreased mortality. In addition, thermoregulation is closely related to the immune response, with inflammatory cytokines playing a key role. Prior studies have assessed the prognostic and immunological implications of static temperature measurements, but there is a paucity of literature studying longitudinal measurements of temperature.
We identified and externally validated four novel subphenotypes of patients with sepsis based on body temperature trajectories. These four trajectory groups had different demographics and comorbidities and had considerable variation in mortality rates, with the “hypothermic” group having a mortality rate three times higher than the “hyperthermic, fast resolvers.” There was also significant variation in levels of commonly tested inflammatory markers between the groups, suggesting that these subphenotypes may represent differential inflammatory responses. Identifying clinically relevant subphenotypes could lead to more personalized sepsis management.
Sepsis is a heterogeneous syndrome characterized by a dysregulated immunological response to infection that results in organ dysfunction and often death (1). Despite the heterogeneity of sepsis, most sepsis trials have focused on a one-size-fits-all approach to treatment, which may explain, in part, the disappointing results of recent studies (2, 3). One of the barriers to effective therapy has been identifying the appropriate subset of patients who would benefit from specific agents (4, 5). Novel methods of identifying subphenotypes among the larger group of patients with sepsis could lead to personalized sepsis management.
An immediately apparent heterogeneity in patients with sepsis is the thermoregulatory response. Although fever is often regarded as the herald of infection, many patients present with normothermia or even hypothermia. Research suggests that body temperature abnormalities carry prognostic information for patients with infection. Several studies have found that hypothermia in the setting of infection is associated with increased mortality, whereas fever is associated with decreased mortality (6–11). In addition to its prognostic utility, temperature may also reveal a patient’s underlying immunological state. The immunological system, specifically inflammatory cytokines, plays a key role in thermoregulation, and prior studies have attempted to correlate pro- and antiinflammatory cytokine levels with temperature abnormalities (8, 12–16). Although these studies have assessed the prognostic and immunological implications of static temperature measurements, it is important to note that temperature is a complex, dynamic variable that changes throughout the natural course of a patient’s illness. In hospitalized patients, body temperature is measured routinely and repeatedly, and each patient’s temperature trajectory provides a repository of longitudinal quantitative data. There is a paucity of research on using this longitudinal data for prognosticating and phenotyping patients with infection.
The aim of this study was to classify infected patients into novel subphenotypes based on their temperature trajectories. We hypothesized that each of these trajectory groups would have different demographics, physiological characteristics, and mortality rates. Temperature trajectories could help identify treatment-responsive subphenotypes in future work and could play a role in realizing the precision medicine approach to sepsis. Some of these results have been reported previously in the form of an abstract (17).
In the model development cohort, all adult patients admitted to University of Chicago Medicine, a 500-bed urban academic tertiary medical center, from November 2008 to January 2016 were eligible for study inclusion in this retrospective analysis of prospectively collected data. Demographic data, International Classification of Diseases, Ninth Revision, billing codes, clinician orders (e.g., blood cultures), vital signs (e.g., tympanic temperature), and laboratory values were collected by the University of Chicago’s Clinical Research Data Warehouse from the electronic health record (Epic). These data were then deidentified and made available on a secure Microsoft SQL server (Microsoft) for analysis. In the validation cohort, all patients admitted to Loyola University Medical Center, a 547-bed urban academic tertiary health system, between 2006 and 2017 were eligible for study inclusion. Data were processed and stored in the same manner as described for the development data. On the basis of general impracticability and minimal harm, waivers of consent were granted by the University of Chicago Institutional Review (IRB#150195) and Loyola University (IRB#209950).
Patients with infection were identified using the criteria published by Rhee and colleagues, which require a blood culture order and at least 4 consecutive days of antibiotics (or antibiotics continued until 1 d before death or discharge), with the first day of antibiotics required to be a parenteral agent given within 48 hours before or after the blood culture order (18). To create a more homogeneous population, the cohort was narrowed to patients who met Rhee criteria for infection in the emergency department and received antibiotics within 24 hours of presentation, with presentation defined as the time of first measured vital sign. Only temperature measurements from Hour 0 (defined as time of first measured vital sign) to Hour 72 were included in the study. The temperature data from Hour 0 to Hour 72 were split into 1-hour blocks of time. If a patient had multiple temperature measures in that 1-hour block, the earliest temperature measurement was used. Nonphysiological temperature data were excluded (temperatures <32°C and >44°C) as per prior publications (19). These criteria were applied in the same manner to both development and validation cohorts.
Demographics, comorbidities, physiological characteristics (i.e., vital signs, laboratory results), and processes of care (e.g., time to antibiotics) were compared between survivors and nonsurvivors in the development cohort. In addition, temperature measurements (mean, maximum, minimum, and presenting temperatures, as well as temperature variability, defined as the SD of temperature measurements during the first 72 h) were compared between survivors and nonsurvivors. For these comparisons, continuous and categorical variables were tested for statistical significance using t tests, Wilcoxon rank-sum tests, or chi-square tests, as appropriate.
In the development cohort, group-based trajectory modeling was used to identify subphenotypes based on temperature trajectories from the first 72 hours of temperature data. Group-based trajectory modeling is a specialized application of finite mixture modeling and is used to identify groups of individuals following similar trajectories for a particular variable of interest (20). The Bayesian information criterion was used for model selection, which penalizes more complex models to ensure that a well-fitting yet parsimonious model is chosen. We tested two-, three-, and four-group models, with both linear and quadratic models investigated, on the basis of prior studies discovering two to four subgroups in sepsis and acute respiratory distress syndrome cohorts (21–23). Group-based trajectory modeling was performed using the traj package in Stata software (StataCorp).
To account for variability in temperature measurement between the development and validation cohorts due to facility-specific factors (e.g., measurement devices), standardization was employed. Specifically, the development cohort’s temperature measurements were standardized on the basis of mean and SD values of this cohort, and a standardized form of the group-based trajectory model was derived. Of note, with group-based trajectory modeling, the trajectory curves and group membership are not altered by using standardized measurements. The temperature measurements in the validation cohort were similarly transformed into standardized values based on this cohort’s mean and SD of temperature measurements. For group assignment in the validation cohort, the standardized temperature measurements from each patient were compared with the predicted measurement of each trajectory group from the model derived in the development cohort at every 1-hour block, and the patient was classified into the group trajectory that resulted in the lowest squared residual error.
Demographics, comorbidities, physiological characteristics, processes of care, and outcomes were compared between the identified trajectory groups with analysis of variance rank tests and chi-square tests. Nonparametric test for trend of clinical characteristics across trajectory groups was performed within the development and validation cohorts, and the significance and directionality of the trends were compared between the two cohorts. Logistic regression was used to determine the association between trajectory group membership and mortality when controlling for age, comorbidities (congestive heart failure [CHF], chronic obstructive pulmonary disease, diabetes, hypertension, liver disease, renal failure, and metastatic cancer), quick Sepsis-related Organ Failure Assessment (qSOFA), and time to antibiotics. Two-tailed P values less than 0.05 were considered statistically significant for all comparisons, and analyses were performed using Stata version 14.1 software.
Of 149,458 unique patient admissions during the study period in the development cohort, 12,413 admissions met criteria for infection in the emergency department with antibiotic administration within 24 hours of presentation and were therefore included in the study (Figure 1). There were 11,673 survivors and 740 nonsurvivors for an overall in-hospital mortality rate of 6% and median length of stay of 4.8 days (IQR, 2.8–8.2 d). The median time to death for nonsurvivors was 5.2 days (IQR, 2.1–10.9 d). There were a total of 264,682 temperature measurements available in the 72 hours from presentation. Mean and median temperatures over this 72-hour period were both 36.5°C (IQR, 36.0–37.1°C), and median qSOFA score was 1 (IQR, 1–2). Over the 72-hours period, fever (temperature, >38°C) occurred in 38% of patients, whereas hypothermia (temperature, <36°C) occurred in 81% of patients. The median number of temperature measurements per patient was 20 measurements. All patients had at least one temperature measurement. Age, comorbidity, and mortality data were available for all patients.
Figure 1. Consolidated Standards of Reporting Trials diagram of patient selection for the study population in the development cohort.
[More] [Minimize]The demographics, comorbidities, and physiological characteristics of survivors and nonsurvivors in the development cohort are shown in Tables 1 and 2. Nonsurvivors were older than survivors, with a median age of 66 years (IQR, 55–78 yr) compared with 59 years (IQR, 45–71 yr). Nonsurvivors had higher rates of CHF, metastatic cancer, and chronic renal failure. Presenting temperature was lower in the nonsurvivors than in survivors (36.3°C vs. 36.9°C; P < 0.001), as were the mean, maximum, and minimum temperatures. Hypothermia was more common in nonsurvivors than in survivors (86% vs. 81%; P < 0.001), whereas fever was less common in nonsurvivors (33% vs. 38%; P < 0.001). For those patients who developed a fever, the time to fever was longer in nonsurvivors than in survivors (22.6 vs. 13.9 h; P < 0.001). Furthermore, temperature variability, as assessed by SD of temperature measurements, was higher in nonsurvivors than in survivors (P < 0.001). The survivors and nonsurvivors followed distinct temperature trajectories: Survivors had a higher presenting temperature followed by a gradual decrease in temperature, whereas nonsurvivors had a lower presenting temperature followed by a gradual increase in temperature (Figure 2).
| Survivors (n = 11,673) | Nonsurvivors (n = 740) | P Value | |
|---|---|---|---|
| Age, yr, median (IQR) | 59 (45–71) | 66 (55–78) | <0.001 |
| Male sex, n (%) | 5,440 (46.6) | 329 (44.5) | 0.3 |
| Race, n (%) | |||
| Black | 8,302 (71.1) | 522 (70.5) | 0.7 |
| White | 2,720 (23.3) | 161 (21.8) | 0.3 |
| Other | 651 (5.6) | 57 (7.7) | 0.02 |
| CHF, n (%) | 2,530 (21.7) | 204 (27.6) | <0.001 |
| COPD, n (%) | 2,393 (20.5) | 124 (16.8) | 0.01 |
| Diabetes mellitus, n (%) | 3,184 (27.3) | 191 (25.8) | 0.4 |
| Hypertension, n (%) | 6,091 (52.2) | 395 (53.4) | 0.5 |
| Metastatic cancer, n (%) | 1,458 (12.5) | 125 (16.9) | <0.001 |
| Liver disease, n (%) | 1,017 (8.7) | 75 (10.1) | 0.2 |
| Chronic renal failure, n (%) | 2,866 (24.6) | 215 (29.1) | 0.006 |
| Survivors (n = 11,673) | Nonsurvivors (n = 740) | P Value | |
|---|---|---|---|
| Presenting temperature, °C | 36.9 | 36.3 | <0.001 |
| Mean temperature, °C | 36.5 | 36.2 | <0.001 |
| Minimum temperature, °C | 35.5 | 34.9 | <0.001 |
| Maximum temperature, °C | 37.8 | 37.6 | <0.001 |
| Temperature, SD | 0.64 | 0.77 | <0.001 |
| Fever (T > 38°C), n (%) | 4,446 (38) | 248 (34) | <0.001 |
| Hypothermia (T < 36°C), n (%) | 9,414 (81) | 639 (86) | <0.001 |
| Time to fever, h, mean (SD) | 14.9 (17.4) | 22.6 (21.5) | <0.001 |
| Mean arterial pressure*, mm Hg, mean (SD) | 91 (19) | 85 (23) | <0.001 |
| Heart rate*, beats/min, mean (SD) | 103 (22) | 107 (27) | <0.001 |
| Respiratory rate*, breaths/min, mean (SD) | 20 (4.5) | 22 (6.9) | <0.001 |
| qSOFA†, mean (SD) | 1.4 (0.5) | 2.0 (0.6) | <0.001 |
| Leukocytosis‡, n (%) | 5,364 (46) | 444 (60) | <0.001 |
| Leukopenia‡, n (%) | 1,186 (10) | 111 (15) | <0.001 |
Figure 2. Body temperature trajectory over time in survivors and nonsurvivors in the development cohort. Trajectory width represents the 95% confidence interval of temperature measurements.
[More] [Minimize]Analysis of the different group-based trajectory models in the development cohort revealed that the four-group quadratic function model had the optimal fit (Figure 3). Group 1 (n = 1,855; 14.9% of the cohort) was characterized by high presenting temperature that was slow to resolve (hereafter denoted as “hyperthermic, slow resolvers”). These patients had the highest mean, minimum, and maximum temperatures (Table 3). Similar to group 1, group 2 (n = 2,877; 23.2%) also presented with a high temperature, but the high temperature decreased at a faster rate (hereafter denoted “hyperthermic, fast resolvers”). Group 3 (n = 4,067; 32.8%) comprised the “normothermic” subjects, with most members (79%) not developing a fever. Finally, group 4 (n = 3,614; 29.1%) had the lowest presenting, mean, minimum, and maximum temperatures, with almost 100% developing hypothermia and only 4% developing a fever (“hypothermic” subjects).
Figure 3. Group-based trajectory model in the development and validation cohorts. Using group-based trajectory modeling, four temperature trajectory group subphenotypes were identified: “hyperthermic, fast resolvers”; “hyperthermic, slow resolvers”; “normothermic”; and “hypothermic.” Temperature measurements were standardized to the hospital’s mean temperature to remove the effect of facility-specific factors (e.g., measurement devices). For the development cohort, standardized temperatures from −1.0 to +1.0 represent temperature measurements from 35.6°C to 37.4°C. For the validation cohort, standardized temperatures from −1.0 to +1.0 represent a range from 36.3°C to 37.7°C.
[More] [Minimize]| All Patients (n = 12,413) | Hyperthermic, Slow Resolvers (n = 1,855) | Hyperthermic, Fast Resolvers (n = 2,877) | Normothermic (n = 4,067) | Hypothermic (n = 3,614) | P Value | |
|---|---|---|---|---|---|---|
| Presenting temperature, °C | 36.8 | 37.8 | 37.8 | 36.5 | 36.0 | <0.001 |
| Mean temperature, °C | 36.5 | 37.5 | 36.7 | 36.6 | 35.9 | <0.001 |
| Minimum temperature, °C | 35.4 | 36.0 | 35.5 | 35.6 | 34.9 | <0.001 |
| Maximum temperature, °C | 37.8 | 39.0 | 38.4 | 37.6 | 36.9 | <0.001 |
| Temperature, SD | 0.65 | 0.81 | 0.81 | 0.56 | 0.55 | <0.001 |
| Fever (T > 38°C), n (%) | 4,694 (38) | 1,758 (95) | 1,938 (67) | 847 (21) | 151 (4) | <0.001 |
| Hypothermia (T < 36°C), n (%) | 10,053 (81) | 725 (39) | 2,484 (86) | 3,236 (80) | 3,608 (100) | <0.001 |
| Time to peak temperature, h, mean (SD) | 19.9 (20.5) | 18.9 (17.8) | 5.8 (9.0) | 26.8 (21.1) | 23.9 (22.2) | <0.001 |
In a sensitivity analysis to assess for informative dropout in the development cohort, patients who died or were discharged before Hour 72 were excluded, and a new trajectory model was tested. This new model classified 95% of the patients into the same groups as the original model. Similarly, in a sensitivity analysis to account for frequency of temperature measurements, a 3-hour time block (instead of a 1-h time block) model was tested, and this model classified 92% of the patients into the same groups as the original model.
In the development cohort, “hyperthermic, slow resolvers” were the youngest patients (median age, 54 yr; IQR, 39–66 yr) and the highest percentage of males. The “hypothermic” group patients were the oldest patients (median age, 63 yr; IQR, 51–76 yr). “Hyperthermic, slow resolvers” had the lowest incidence of all comorbidities tested; 43% of patients had no comorbidities. “Hypothermic” patients had the highest incidence of comorbidities, with 71% of the group having one or more comorbidities (see Table E1 in the online supplement).
In the development cohort, as shown in Table E2, “hyperthermic, slow resolvers” had the highest incidence of leukopenia and leukocytosis, as well as the highest erythrocyte sedimentation rate (ESR) and C-reactive protein (CRP) levels. “Hyperthermic, slow resolvers” also had the lowest creatinine, greatest urine output per day, and highest heart rate. Together with the “hyperthermic, fast resolvers,” they had the lowest lactic acid levels. “Hypothermic” patients had the lowest incidence of leukopenia and leukocytosis, the lowest ESR and CRP levels, and the highest creatinine and lactic acid levels. They also had the lowest urine output and slowest heart rate. There were no differences of clinical significance detected between groups for mean arterial pressure and qSOFA scores.
In the development cohort, as shown in Table E3, “hyperthermic, fast resolvers” had the shortest time to antibiotics and the lowest percentage requiring vasopressors (4.6%). “Hyperthermic, slow resolvers” had the highest exposure to acetaminophen and the lowest exposure to both prednisone and methylprednisolone over the 72-hour period. “Hypothermic” patients had the highest percentage requiring vasopressors (8.4%), lowest exposure to acetaminophen, and most exposure to both prednisone and methylprednisolone. Acetaminophen administration differences between the four trajectory groups are presented in Figure E1. In-hospital mortality was highest in the “hypothermic” group (9.5%), followed by the “normothermic” and “hyperthermic, slow resolvers” (5.3% and 5.1%, respectively), with the “hyperthermic, fast resolvers” having the lowest mortality (2.9%).
In logistic regression in the development cohort, when controlling for age, comorbidities (CHF, chronic obstructive pulmonary disease, diabetes, hypertension, liver disease, renal failure, and metastatic cancer), qSOFA, and time to antibiotics, membership in the “hyperthermic, fast resolvers” group was associated with decreased mortality risk (odds ratio [OR], 0.57; 95% confidence interval [CI], 0.44–0.75; P < 0.001) compared with the “normothermic” group. Membership in the “hypothermic” group was associated with increased mortality risk (OR, 1.68, 95% CI, 1.40–2.03; P < 0.001) compared with the “normothermic” group.
In the validation dataset, 19,053 patients met criteria for infection in the emergency room and received antibiotics within 24 hours. There were 17,892 survivors and 1,161 nonsurvivors for an in-hospital mortality rate of 6.1% and median length of stay of 3.9 days (IQR, 2.1–7.0 d). Application of the derived model to the validation cohort resulted in the following membership distribution: “hyperthermic, slow resolvers” (n = 2,087; 11.0%), “hyperthermic, fast resolvers” (n = 4,207; 22.1%), “normothermic” (n = 7,278; 38.2%), and “hypothermic” (n = 5,481; 28.8%) (Figure 3).
Although the overall mean and median ages were higher in the validation cohort, the relative trend in age distributions between the groups was similar to the development cohort (Table 4). The “hyperthermic, slow resolvers” was the youngest group, with a median age of 57 years (IQR, 43–68 yr), and the “hypothermic” group was the oldest, with a median age of 68 years (IQR, 54–80 yr). The relative distribution of comorbidities among the four groups was also similar to that of the development cohort. The “hyperthermic, slow resolvers” had the lowest prevalence of CHF, chronic obstructive pulmonary disease, hypertension, liver disease, and chronic renal failure. The “hypothermic” group had the highest prevalence of all comorbidities.
| Development Cohort | Validation Cohort | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| HSR | HFR | NT | HT | P Value* | HSR | HFR | NT | HT | P Value* | |
| Membership, n | 1,855 | 2,877 | 4,067 | 3,614 | — | 2,087 | 4,207 | 7,278 | 5,481 | — |
| Age, yr, median | 54 | 56 | 60 | 63 | <0.001 | 57 | 61 | 62 | 68 | <0.001 |
| Male sex, % | 50 | 48 | 45 | 46 | 0.001 | 53 | 50 | 48 | 50 | 0.01 |
| CHF, % | 13 | 20 | 22 | 29 | <0.001 | 12 | 13 | 14 | 21 | <0.001 |
| COPD, % | 15 | 21 | 20 | 23 | <0.001 | 13 | 15 | 16 | 19 | <0.001 |
| Liver disease, % | 6.6 | 8.4 | 8.8 | 10 | <0.001 | 6.9 | 7.3 | 7.3 | 8.8 | 0.004 |
| Metastatic cancer, % | 11 | 13 | 13 | 13 | 0.2 | 6.4 | 5.4 | 5.9 | 6.7 | 0.1 |
| Chronic renal failure, % | 18 | 26 | 24 | 29 | <0.001 | 16 | 19 | 18 | 22 | <0.001 |
| qSOFA† | 1.5 | 1.4 | 1.4 | 1.5 | 0.06 | 1.4 | 1.4 | 1.3 | 1.4 | 0.01 |
| ESR, mm/h | 77 | 68 | 71 | 66 | <0.001 | 70 | 66 | 63 | 53 | <0.001 |
| CRP, mg/L | 173 | 125 | 106 | 84 | <0.001 | 198 | 165 | 131 | 99 | <0.001 |
| Leukocytosis‡, % | 52 | 48 | 48 | 42 | <0.001 | 61 | 62 | 59 | 56 | <0.001 |
| Acetaminophen, % | 90 | 75 | 50 | 41 | <0.001 | 95 | 81 | 59 | 50 | <0.001 |
| Methylprednisolone, % | 3.1 | 7.6 | 8.1 | 9.2 | <0.001 | 4.9 | 6.5 | 7.0 | 11 | <0.001 |
| Mortality, % | 5.1 | 2.9 | 5.3 | 9.5 | <0.001 | 10.2 | 3.0 | 4.5 | 9.0 | <0.001 |
Similar to the development cohort, the ESR and CRP levels were highest in the “hyperthermic, slow resolvers” and lowest in the “hypothermic” group (Table 4). In addition, the “hyperthermic, slow resolvers” had the lowest exposure to corticosteroids, whereas the “hypothermic” group had the highest exposure. Acetaminophen administration decreased across groups from “hyperthermic, slow resolvers” to “hyperthermic, fast resolvers” to “normothermic” to “hypothermic.” The remainder of the trajectory group characteristics in the validation cohort are detailed in the online supplement (Tables E4–E6).
In the validation cohort, in-hospital mortality was highest in “hyperthermic, slow resolvers” (10.2%), followed by “hypothermic” (9.0%), then “normothermic” (4.5%), and lowest in “hyperthermic, fast resolvers” (3%). In logistic regression, membership in the “hyperthermic, fast resolvers” group was associated with lower mortality risk (OR, 0.55; 95% CI, 0.44–0.68; P < 0.001), whereas membership in the “hypothermic” group was associated with higher mortality risk (OR, 1.68; 95% CI, 1.44–1.96; P < 0.001) than in the “normothermic” group. The “hyperthermic, slow resolvers” group was also associated with higher mortality risk (OR, 2.15; 95% CI, 1.77–2.61; P < 0.001) than in the “normothermic” group.
We report a novel method to identify subphenotypes in patients with sepsis on the basis of temperature trajectories. Through group-based trajectory modeling, we discovered and validated four groups of patients that follow distinct temperature trajectories. We found significant demographic and physiological differences between the trajectory groups, as well as considerable variation in mortality, with the “hypothermic” group having a mortality rate three times higher than the “hyperthermic, fast resolvers.” Finally, we found that the levels of inflammatory markers, such as ESR and CRP, and incidence of leukocytosis were distinctly varied between the trajectory groups. These findings have important implications for understanding the heterogeneity of patients with sepsis and may inform future work to determine treatment response across different subphenotypes.
Recent studies provide evidence that the identification of subphenotypes could lead to more personalized management strategies. For example, Calfee and colleagues identified subphenotypes with different inflammatory responses in acute respiratory distress syndrome using physiological and biomarker data, with the subphenotypes displaying different clinical outcomes with high versus low positive end-expiratory pressure strategies (21). Similarly, Wong and colleagues identified endotypes in pediatric and adult sepsis using gene expression signatures, with the finding that adjunctive corticosteroids were associated with increased mortality in one endotype but not the other in the pediatric population (22, 24). The universal availability of longitudinal temperature trajectories makes it a readily accessible adjunct to more resource-intensive diagnostics for the identification of clinically relevant sepsis subphenotypes.
Similarly to prior research, we found that fever was more common in survivors, whereas hypothermia was more common in nonsurvivors (6–11). Young and colleagues have previously reported that higher peak temperatures were associated with a reduced in-hospital mortality rate in patients with infection, whereas hypothermia was associated with an increased mortality rate (11). Notably, we discovered that transient hypothermia (temperature, <36°C) was very common, with 81% of patients in the development cohort having at least one episode in the 72 hours from presentation. We also report that temperature variability was higher in nonsurvivors than in survivors. Prior studies have also reported increased SD and variability of temperature measurements in nonsurvivors compared with survivors with sepsis; our study confirms these findings in a much larger patient population both inside and outside ICUs (25–27). These data support the prognostic utility of dynamic temperature measurements and their benefit over using a single temperature measurement.
Importantly, we discovered and validated four temperature trajectory groups: “hyperthermic, slow resolvers”; “hyperthermic, fast resolvers”; “normothermic”; and “hypothermic.” On the basis of our results, we hypothesize that “hyperthermic, slow resolvers” represent a hyperinflammatory subphenotype, because the high incidence of leukocytosis and elevated ESR and CRP in this group provides evidence of a hyperinflammatory state. In addition, these patients were the youngest and had the least comorbidities and were therefore less likely to be immunosuppressed. Finally, the lack of exposure to corticosteroids may have played a role in maintaining a hyperinflammatory state. In contrast, we hypothesize that the “hypothermic” group represents a hypoinflammatory subphenotype. The “hypothermic” subjects were the oldest and had the most comorbidities and were therefore more likely to be immunosuppressed (28). The lower incidence of leukocytosis and lower ESR and CRP provide additional evidence of a hypoinflammatory state. Furthermore, the greater exposure to corticosteroids may have contributed to a suppressed inflammatory response. Given the low mortality rate of the “hyperthermic, fast resolvers,” we hypothesize that they may represent a well-balanced inflammatory subphenotype. Specifically, these patients may launch an early inflammatory response and appropriately attenuate this response once the pathogen is cleared, thereby avoiding inflammatory injury.
Studies have revealed that sepsis is characterized by both pro- and antiinflammatory responses, often concomitantly; however, the precise immunological characterization of individual patients remains elusive (29–33). With the growing interest in immunomodulatory therapy, accurate immunological phenotyping of patients with sepsis is essential (33–35). Thermoregulation is closely linked to the immunological system, and thus temperature patterns may play a role in immunological phenotyping (12). Tumor necrosis factor-α, IL-1β, and IL-6 are mediators underlying the febrile response, whereas there is evidence that IL-10 may play a role in fever defervescence (13, 36). If temperature trajectory groups are found to have unique cytokine signatures, these groups would be a readily accessible way to characterize patients with sepsis, and this could lead to targeted therapy. Even without specific cytokine signatures, temperature trajectory groups could be used to differentiate hyper- and hypoinflammatory states to guide the decision between immunosuppressive and immunostimulatory therapy in sepsis. For example, corticosteroids or other immunosuppressive therapy could be directed to hyperinflammatory patients, whereas granulocyte-macrophage colony-stimulating factor or anti-PDL1 could be directed to hypoinflammatory patients.
Our study has several limitations. First, we do not have immunological markers such as cytokine levels to confirm the hypothesized differential immunological basis for these subphenotypes. Further exploration of the association of trajectory groups and immunological markers is necessary. Second, we did not control for the site of infection, pathogen, or adequacy of antibiotics, all of which may contribute to differential temperature trajectories. Third, we could not remove the effect of medications, such as acetaminophen, on temperature trajectories. However, on the basis of pattern of acetaminophen administration in our study (Figure E1) as well as studies such as the HEAT (Permissive Hyperthermia through Avoidance of Acetaminophen in Known or Suspected Infection in the Intensive Care Unit) trial, we do not believe that acetaminophen administration shaped temperature trajectories in a significant way (37). Further research into the role of pathogens, location of infection, and medications such as acetaminophen in temperature trajectory is necessary to determine how these factors impact the subphenotypes.
In conclusion, we have discovered four novel temperature trajectory subphenotypes of infected patients, and this approach provides a framework for prognosticating and identifying distinct subphenotypes of patients with sepsis. Our results suggest that the four groups may represent differential immunological responses to infection. If confirmed to relate to the underlying immunological response, temperature trajectory groups could identify treatment-responsive subphenotypes and lead to improved personalization of sepsis management.
The authors thank Timothy Holper, M.S., Julie Johnson, M.P.H., R.N., and Thomas Sutton, M.S., for assistance with data abstraction; Laura Ruth Venable, M.S., and Mary Akel, M.P.H., for the administrative support; and Dana Edelson, M.D., M.S., for helpful guidance during project conception.
| 1. | Singer M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA 2016;315:801–810. |
| 2. | Mullard A. Drug withdrawal sends critical care specialists back to basics. Lancet 2011;378:1769. |
| 3. | Angus DC. The search for effective therapy for sepsis: back to the drawing board? JAMA 2011;306:2614–2615. |
| 4. | Cohen J, Opal S, Calandra T. Sepsis studies need new direction. Lancet Infect Dis 2012;12:503–505. |
| 5. | Cohen J, Vincent JL, Adhikari NK, Machado FR, Angus DC, Calandra T, et al. Sepsis: a roadmap for future research. Lancet Infect Dis 2015;15:581–614. |
| 6. | Clemmer TP, Fisher CJ Jr, Bone RC, Slotman GJ, Metz CA, Thomas FO; The Methylprednisolone Severe Sepsis Study Group. Hypothermia in the sepsis syndrome and clinical outcome. Crit Care Med 1992;20:1395–1401. |
| 7. | Drewry AM, Fuller BM, Skrupky LP, Hotchkiss RS. The presence of hypothermia within 24 hours of sepsis diagnosis predicts persistent lymphopenia. Crit Care Med 2015;43:1165–1169. |
| 8. | Arons MM, Wheeler AP, Bernard GR, Christman BW, Russell JA, Schein R, et al.; Ibuprofen in Sepsis Study Group. Effects of ibuprofen on the physiology and survival of hypothermic sepsis. Crit Care Med 1999;27:699–707. |
| 9. | Rumbus Z, Matics R, Hegyi P, Zsiboras C, Szabo I, Illes A, et al. Fever is associated with reduced, hypothermia with increased mortality in septic patients: a meta-analysis of clinical trials. PLoS One 2017;12:e0170152. |
| 10. | Peres Bota D, Lopes Ferreira F, Mélot C, Vincent JL. Body temperature alterations in the critically ill. Intensive Care Med 2004;30:811–816. |
| 11. | Young PJ, Saxena M, Beasley R, Bellomo R, Bailey M, Pilcher D, et al. Early peak temperature and mortality in critically ill patients with or without infection. Intensive Care Med 2012;38:437–444. |
| 12. | Rehman T, deBoisblanc BP. Persistent fever in the ICU. Chest 2014;145:158–165. |
| 13. | Dinarello CA. Cytokines as endogenous pyrogens. J Infect Dis 1999;179:S294–S304. |
| 14. | van Dissel JT, van Langevelde P, Westendorp RG, Kwappenberg K, Frölich M. Anti-inflammatory cytokine profile and mortality in febrile patients. Lancet 1998;351:950–953. |
| 15. | Wiewel MA, Harmon MB, van Vught LA, Scicluna BP, Hoogendijk AJ, Horn J, et al. Risk factors, host response and outcome of hypothermic sepsis. Crit Care 2016;20:328. |
| 16. | Marik PE, Zaloga GP; Norasept II Study Investigators. Hypothermia and cytokines in septic shock. Intensive Care Med 2000;26:716–721. |
| 17. | Bhavani S, Carey K, Edelson D, Churpek MM. Association between temperature variability and mortality in hospitalized patients with infection [abstract]. Am J Respir Crit Care Med 2018;197:A6060. |
| 18. | Rhee C, Kadri S, Huang SS, Murphy MV, Li L, Platt R, et al. Objective sepsis surveillance using electronic clinical data. Infect Control Hosp Epidemiol 2016;37:163–171. |
| 19. | Churpek MM, Zadravecz FJ, Winslow C, Howell MD, Edelson DP. Incidence and prognostic value of the systemic inflammatory response syndrome and organ dysfunctions in ward patients. Am J Respir Crit Care Med 2015;192:958–964. |
| 20. | Nagin DS, Odgers CL. Group-based trajectory modeling in clinical research. Annu Rev Clin Psychol 2010;6:109–138. |
| 21. | Calfee CS, Delucchi K, Parsons PE, Thompson BT, Ware LB, Matthay MA; NHLBI ARDS Network. Subphenotypes in acute respiratory distress syndrome: latent class analysis of data from two randomised controlled trials. Lancet Respir Med 2014;2:611–620. |
| 22. | Wong HR, Cvijanovich NZ, Anas N, Allen GL, Thomas NJ, Bigham MT, et al. Developing a clinically feasible personalized medicine approach to pediatric septic shock. Am J Respir Crit Care Med 2015;191:309–315. |
| 23. | Scicluna BP, van Vught LA, Zwinderman AH, Wiewel MA, Davenport EE, Burnham KL, et al.; MARS Consortium. Classification of patients with sepsis according to blood genomic endotype: a prospective cohort study. Lancet Respir Med 2017;5:816–826. |
| 24. | Wong HR, Sweeney TE, Hart KW, Khatri P, Lindsell CJ. Pediatric sepsis endotypes among adults with sepsis. Crit Care Med 2017;45:e1289–e1291. |
| 25. | Varela M, Calvo M, Chana M, Gomez-Mestre I, Asensio R, Galdos P. Clinical implications of temperature curve complexity in critically ill patients. Crit Care Med 2005;33:2764–2771. |
| 26. | Varela M, Jimenez L, Fariña R. Complexity analysis of the temperature curve: new information from body temperature. Eur J Appl Physiol 2003;89:230–237. |
| 27. | Drewry AM, Fuller BM, Bailey TC, Hotchkiss RS. Body temperature patterns as a predictor of hospital-acquired sepsis in afebrile adult intensive care unit patients: a case-control study. Crit Care 2013;17:R200. |
| 28. | Opal SM, Girard TD, Ely EW. The immunopathogenesis of sepsis in elderly patients. Clin Infect Dis 2005;41(Suppl 7):S504–S512. |
| 29. | Boomer JS, To K, Chang KC, Takasu O, Osborne DF, Walton AH, et al. Immunosuppression in patients who die of sepsis and multiple organ failure. JAMA 2011;306:2594–2605. |
| 30. | Novotny AR, Reim D, Assfalg V, Altmayr F, Friess HM, Emmanuel K, et al. Mixed antagonist response and sepsis severity-dependent dysbalance of pro- and anti-inflammatory responses at the onset of postoperative sepsis. Immunobiology 2012;217:616–621. |
| 31. | Osuchowski MF, Welch K, Siddiqui J, Remick DG. Circulating cytokine/inhibitor profiles reshape the understanding of the SIRS/CARS continuum in sepsis and predict mortality. J Immunol 2006;177:1967–1974. |
| 32. | Osuchowski MF, Welch K, Yang H, Siddiqui J, Remick DG. Chronic sepsis mortality characterized by an individualized inflammatory response. J Immunol 2007;179:623–630. |
| 33. | 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. |
| 34. | Meisel C, Schefold JC, Pschowski R, Baumann T, Hetzger K, Gregor J, et al. Granulocyte-macrophage colony-stimulating factor to reverse sepsis-associated immunosuppression: a double-blind, randomized, placebo-controlled multicenter trial. Am J Respir Crit Care Med 2009;180:640–648. |
| 35. | Osuchowski MF, Connett J, Welch K, Granger J, Remick DG. Stratification is the key: inflammatory biomarkers accurately direct immunomodulatory therapy in experimental sepsis. Crit Care Med 2009;37:1567–1573. |
| 36. | Cartmell T, Ball C, Bristow AF, Mitchell D, Poole S. Endogenous interleukin-10 is required for the defervescence of fever evoked by local lipopolysaccharide-induced and Staphylococcus aureus-induced inflammation in rats. J Physiol 2003;549:653–664. |
| 37. | Young P, Saxena M, Bellomo R, Freebairn R, Hammond N, van Haren F, et al.; HEAT Investigators; Australian and New Zealand Intensive Care Society Clinical Trials Group. Acetaminophen for fever in critically ill patients with suspected infection. N Engl J Med 2015;373:2215–2224. |
Supported by a career development award from the NHLBI (K08 HL121080) and an R01 grant from the National Institute of General Medical Sciences (R01 GM123193) (M.M.C.) and by the National Center for Advancing Translational Science of the NIH under award ULITR002389 (P.A.V.). M.M.C. has a patent pending (ARCD. P0535 US.P2) for risk stratification algorithms for hospitalized patients.
Author Contributions: Study concept and design: S.V.B., P.A.V., and M.M.C.; acquisition of data: E.R.G., M.A., and M.M.C.; analysis and interpretation of data: all authors; first draft of the manuscript: S.V.B.; critical revision of the manuscript for important intellectual content: all authors; statistical analysis: S.V.B. and K.A.C.; obtaining funding: P.A.V. and M.M.C.; administrative, technical, and material support: K.A.C., E.R.G., M.A., and M.M.C.; study supervision: P.A.V. and M.M.C.; data access and responsibility: S.V.B. All authors had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.
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.201806-1197OC on February 21, 2019
Author disclosures are available with the text of this article at www.atsjournals.org.
