Our study was conducted retrospectively through the eICU Collaborative Research Database (eICU-CRD),13 an online database that contains over 200,000 ICU admissions and is monitored by multiple centers throughout the continental United States. From 2014 to 2015, all data were recorded automatically and accessed electronically using the Philips Healthcare eICU software.13

The eICU-CRD has been utilized in prior observational research.14–16 We completed the Collaborative Institutional Training Initiative (CITI) program and obtained certification in compliance with the data usage agreement set by the PhysioNet Review Board. Access to the database aligns with the Safe Harbor provision of the Health Insurance Portability and Accountability Act (HIPAA). Data access was granted based on completion of the CITI program, specifically ‘Data or Specimens Only Research’. Given the use of a publicly accessible database without patient involvement, we obtained certification from Privacert (Cambridge, MA) to ensure compliance with safe harbor standards regarding re-identification risk. As a result, the investigation did not require endorsement from the Institutional Review Board of the Massachusetts Institute of Technology (record ID 47549485) or the acquisition of informed consent. The study was conducted in accordance with the principles of the Declaration of Helsinki, adhering to all relevant rules and regulations.

The subjects were all patients diagnosed with sepsis at the time of admission to ICU.

Sepsis was defined as suspected or confirmed infection, with a Sequential Organ Failure Assessment (SOFA) score exceeding 2 points in the Acute Physiology and Chronic Health Evaluation (APACHE) IV dataset.1,17 The eICU Collaborative Research Database contains ICD-9 codes that can be used to indicate infections.

We applied the subsequent exclusion criteria: (1) Patients who were not admitted to the ICU for the first time; (2) missing temperature after ICU admission; (3) temperature < 30 °C; (4) age < 18 years old; (5) lacking of hospitalization outcome. Because it was a secondary analysis of the database, our study did not calculate the sample size. The study flowchart was presented in Fig. 1.

Figure 1.

The study flowchart.

(0.25MB).

The eICU database contains demographic information, physiological measurements from bedside monitors, diagnoses using the International Classification of Diseases, 9th Edition, Clinical Modification (ICD-9-CM) codes, severity of illness assessments, laboratory results, and treatment details.

Data was collected from the eICU-CRD for all participants within the first 24 hours after admission. Extracting baseline patient characteristics such as age, sex, race and weight from the patient table and the Apache Patient Result table; The physiological variables and treatment information of patients were obtained from Apache Aps Var table: Body temperature (°C), Respiratory Rate (RR), Heart Rate (HR), Mean Arterial Pressure (MAP), Mechanical ventilation use, Vasopressor use (1st 24 h) and Hemodialysis, laboratory indicators such as Initial lactate level, Creatinine, and White Blood Cells (WBC) were collected from the laboratory tables; Comorbidities including Metastatic cancer, Immunosuppression, Acute Myocardial Infarction (AMI), Arrhythmia, Congestive Heart Failure(CHF), Hepatic Failure, Diabetes, Chronic Obstructive Pulmonary Disease(COPD) were extracted from the APACHE IV score. In this study, the collected body temperature data represent the highest temperatures recorded within the first 24 hours after patients were admitted to the Intensive Care Unit (ICU), primarily obtained through oral and rectal thermometry. Additionally, we identified diagnostic codes for sepsis from the diagnostic form, measuring disease severity by SOFA score, APACHEIV score and Acute Physiology score III. To address potential bias stemming from missing covariates in the modeling process, this study adopts multiple imputation techniques for managing absent data. This strategy aims to enhance the precision of statistical analysis on the intended sample.18,19

The study investigated all-cause mortality occurring within 28 days following admission to the Intensive Care Unit (ICU).

We used means ± standard deviation (SD) or median and interquartile ranges (IQR) to represent continuous variables, while categorical data were presented using counts and percentages. To analyze differences among body temperature tertiles, we utilized one-way analysis of variance (ANOVA) for continuous variables and chi-squared tests for categorical variables (Table 1). Furthermore, unadjusted correlations between baseline metrics and 28-day mortality were also compared (Table 2).

Univariate and multivariate binary logistic regression analyses were performed, and three models were constructed to evaluate the association between body temperature and 28-day mortality: model 1 did not consider any covariates, model 2 adjusted for sociodemographic data, and model 3 included the covariates from model 2 as well as other confounding factors (as shown in Table 3). The adjustment of other covariates was guided by clinical expertise, literature findings, and the outcomes of univariate analysis (detailed in Table 2): Source of infection, WBC, creatinine, Initial lactate level, Respiratory Rate (bpm), Heart Rate (bpm), MAP (mmHg), SOFA score, Mechanical ventilation use, Dialysis, Vasopressor use (1st 24 h), Metastatic cancer, Arrhythmia, CHF, Hepatic failure, COPD.

Given that logistic regression is unable to account for nonlinear relationships, we employed a penalized spline method for smooth curve fitting to account for the possibility of a nonlinear association between body temperature and 28-day mortality, as shown in Fig. 2. In cases where nonlinearity was identified, we used a recursive algorithm to estimate the inflection point, followed by a bootstrapping algorithm to determine the range of the inflection point and calculate its confidence interval (CI). Subsequently, we developed a two-phase linear regression model on either side of the inflection point.20,21 We determined the best-fit model (linear regression model vs. two-phase linear regression model) based on the p-values obtained from the log likelihood ratio test (Table 4). We employed this approach to account for the possibility of a nonlinear association between body temperature and 28-day mortality.

Figure 2.

Associations between body temperature and 28-day mortality in all patients with sepsis. A threshold, nonlinear association between the body temperature and 28-day mortality was found in a generalized additive model (GAM). Solid rad line represents the smooth curve fit between variables. Blue bands represent the 95% of confidence interval from the fit. Adjusted for Admission weight; Ethnicity; WBC; Creatinine; Initial lactate level; Respiratory rate; Heart rate; MAP; SOFA; Mechanical ventilation use; Dialysis; Vasopressor use (1st 24 h); Metastatic cancer; Arrhythmia; CHF; Hepatic failure; COPD; Source of infection.

(0.16MB).

In our study, the number of participants with missing data of Admission weight, Ethnicity, WBC, Creatinine, Initial lactate level (mmol/L), Respiratory rate (bpm), Heart rate (bpm), MAP (mmHg), SOFA, APACHE IV score, Acute Physiology Score III was 374(2.07%), 3254(18.02%), 2185(12.10%), 45(0.25%), 6815(37.74%), 53(0.29%, 16(0.09%), 40(0.22%), 532(2.95%), 2041(11.30%) respectively. To prevent a decrease in statistical test power and bias resulting from the direct exclusion of missing values, we employed multiple imputation using chained equations (MICE) based on SAS to estimate the missing data. The imputed data were then analyzed,18,19 and the interpolated data had little effect on the outcome. The analyses were performed with the statistical software packages R (http://www.R-project.org, The R Foundation) and Empower Stats (http://www.empowerstats.com, X&Y Solutions, Inc, Boston, MA). P values less than 0.05 (two-sided) were considered statistically significant.

Analyzed in this study were data from 18,056 patients, with an average age of 65.82 ± 16.23 years, and nearly half of them (50.94%) were male. Table 1 illustrates a comparison of various aspects among patients in different body temperature tertiles, including demographics, vital signs, laboratory test results, site of infection, treatment details, and severity of illness. The table indicates that compared to the highest temperature group, the hypothermia group had older and lighter patients upon admission, elevated levels of creatinine and lactate, higher SOFA scores, APACHE IV scores, acute physiology scores, and longer ICU stays. Notably, the hypothermia group exhibited the highest mortality rate among the three groups, reaching 15.71%.

The results of the univariate logistic regression analysis presented in Table 2 indicate several key associations between baseline variables and 28-day mortality among the 18,056 sepsis patients. Significant positive correlations were observed with initial lactate level, creatinine, heart rate, respiratory rate, advanced age (>65 years), SOFA score, APS III score, and APACHE IV score (all p < 0.0001). Notably, metastatic cancer, acute myocardial infarction, arrhythmia, chronic heart failure, hepatic failure, mechanical ventilation use, and vasopressor use within the first 24 hours were also associated with increased mortality risk. Conversely, body temperature and diabetes were negatively associated with 28-day mortality (p < 0.0001). No significant associations were found with gender, ethnicity, immunosuppression, COPD, or dialysis.

We utilized the binary logistic regression model to construct three models for the purpose of examining the potential association between body temperature and the 28-day mortality rate of sepsis patients, the effect sizes and 95% CIs are listed in Table 3. In Model 1, an unadjusted model, hypothermia (T2: 30.00–36.28 °C) was significantly associated with an increased risk of 28-day mortality (OR = 2.18, 95% CI: 1.93–2.47, p < 0.0001), indicating that hypothermia is positively associated with mortality. After adjusting for socio-demographic variables (age, weight, ethnicity) in Model 2, the results remained consistent, with hypothermia showing a positive association with 28-day mortality (OR = 2.17, 95% CI: 1.92–2.46, p < 0.0001). In the fully-adjusted model (Model 3), which accounted for additional covariates listed in Table 2, hypothermia continued to be positively associated with 28-day mortality (OR = 1.57, 95% CI: 1.37–1.80, p < 0.0001). Conversely, the hyperthermia group (T3: 36.70–41.90 °C) did not show a statistically significant association with 28-day in-hospital mortality across any of the models, including the fully-adjusted Model 3 (OR = 1.07, 95% CI: 0.93–1.23, p = 0.3464).

A nonlinear dose-response relationship between body temperature and 28-day ICU in-hospital mortality was observed using a generalized summation model and curve fitting (Fig. 3). The inflection point for body temperature was determined to be 36.67 °C using a recursive algorithm. To account for this nonlinear relationship, a two-piecewise binary logistic regression model was developed on either side of the inflection point. The model was compared to a standard linear regression model, and the p-value of the log-likelihood ratio test indicated a significantly better fit for the two-piecewise model (p < 0.0001; Table 4). As shown in Table 4, when body temperature was ≤36.67 °C, each 1 °C decrease in body temperature was associated with a significant increase in 28-day in-hospital mortality (OR = 0.74, 95% CI: 0.70–0.80, p < 0.0001), indicating that lower body temperatures are linked to higher mortality risk. In contrast, for patients with body temperature >36.67 °C, there was no statistically significant association between temperature and mortality (OR = 1.07, 95% CI: 1.00–1.15, p = 0.0606), suggesting that hyperthermia might not have a substantial impact on 28-day in-hospital mortality among sepsis patients.

Figure 3.

The distribution of body temperature in the total population.

(0.13MB).