Determinants of long-term outcome in ICU survivors: results from the FROG-ICU study

FROG-ICU was a prospective, observational, multicenter cohort study in which survivors of critical illness were followed up for up to 1 year post ICU discharge. The study was conducted in France and Belgium in accordance with Good Clinical Practice (Declaration of Helsinki 2002) and Ethical Committee approvals (Comité de Protection des Personnes—Ile de France IV, IRB n°00003835 and Commission d’éthique biomédicale hospitalo-facultaire de l’hôpital de Louvain, IRB n°B403201213352). It is registered on ClinicalTrials.gov (NCT01367093). Patients were included from August 2011 to June 2013. Details of design and methods have been published previously [ 14]. All patients admitted to any of the participating centers during the recruitment period who met the eligibility criteria and survived their ICU stay had a clinical examination and biological tests performed at discharge from the ICU, and were followed up for 1 year through telephone calls and postal questionnaires at 3, 6 and 12 months.

The study involved 21 medical, surgical or mixed ICUs in 14 university hospitals. Inclusion criteria were: invasive mechanical ventilation support for at least 24 h and/or treatment with a vasoactive agent (except dopamine) for more than 24 h. Noninclusion criteria were: age younger than 18 years old; severe head injury (initial Glasgow Coma Scale ≤ 8), brain death or a persistent vegetative state; pregnancy or breastfeeding; transplantation in the past 12 months; moribund patient; and/or no social security coverage. The Ethical Committees waived the need for written consent; all patients and/or next of kin were informed and oral consent was documented in the patients’ medical records by the investigator.

The primary purpose of the FROG-ICU study was to assess the incidence of all-cause mortality in the year following ICU discharge, and to identify independent factors associated with mortality. The main secondary objective of FROG-ICU was to evaluate the association between circulation cardiovascular biomarkers levels at discharge and 1-year mortality.

Details of data collection have been reported previously [ 14]. Briefly, clinical and biological data were recorded at admission, during the ICU stay and at discharge from the ICU. In order to explore the mechanisms of clinical abnormalities at ICU discharge associated with subsequent deaths, cardiovascular biomarkers were collected at discharge and measured centrally. These included markers of: cardiac failure (NT-proBNP; Roche Diagnostics GmbH, Mannheim, Germany); cardiac ischemia (hs-TnI; Abbott, Abbott Park, IL, USA); vascular dysfunction [ 15] (bio-ADM; Adrenomed GmbH, Hennigsdorf, Germany); and cardiac stress (sST2; Eurobio, Critical Diagnostics, San Diego, CA, USA) which prognosticates for cardiovascular death [ 16, 17]. The deprivation index (FDep) was used as a measure of socioeconomic inequalities in health status. The FDep is based on the patients’ residential zip codes and was specifically developed for the French context using the following four variables to compute a single composite index: median household income, percentage of high school graduates in the population aged ≥ 15 years, percentage of blue-collar workers in the active population and unemployment rate [ 18].

Results are expressed as median (interquartile range (IQR)) or count (percentage) as appropriate. The primary analysis examining factors associated with 1-year mortality was based on analysis of the clinical and biological variables measured in patients discharged alive from the ICU. Marginal associations between single variables and 1-year mortality were assessed by a Wilcoxon rank-sum test for quantitative variables and the chi-square test for qualitative variables. Multivariable logistic regression was used to determine a set of variables independently associated with 1-year mortality. Variables associated with outcome at a 0.05 level and with less than 20% of missing data were considered within the multivariable model. The log-linearity of the quantitative variables was evaluated systematically, and, if appropriate, variable transformation was performed. Log-linearity of the association between continuous variables and the outcome was checked using a cubic spline and the Wald test. Cutoff values were derived from the plots of the effect according to the value of the variable of interest. Missing values were handled by multiple imputation by chained equations (MICE) [ 19]. All variables selected for the multivariable model were considered in the imputation model. A total of 51 imputed samples was generated using 15 iterations of the chained equation process. A selection model process was performed using a backward stepwise approach with stopping rules based on a cutoff at 0.05 for p values. At each step of the selection, inference was combined from the sets of imputed samples using Rubin’s rules [ 20]. The existence of any colinearities was observed, and a test of goodness of fit was performed using the Hosmer–Lemeshow test on the complete case model [ 21]. Measures of association consisted of odds ratios (ORs) and their confidence intervals (CIs) at 95% estimated using Rubin’s rules. The predictive power of the four biomarkers of interest was assessed using receiver operating curve (ROC) analyses. The area under the ROC (AUC) was estimated for each biomarker. For both the clinical model and the clinical model including biomarker information, the AUCs were estimated from the sets of imputed samples using Rubin’s rules. The latter two were compared using the Delong test. As it is now recognized that highlighting a statistically significant association between new biomarkers and patient outcomes is not sufficient to demonstrate the interest of these biomarkers in terms of risk prediction [ 22– 24], we used the proposed methodology of Pencina et al. [ 23], which has been used in multiple articles of application. The net reclassification improvement (NRI) and integrated discrimination improvement (IDI) of each biomarker added to the full clinical model will be calculated, and comparisons between different biomarkers will be performed [ 23].

Calculating the number of subjects required was based on the primary endpoint; that is, the risk factors associated with 1-year all-cause mortality. Study of the literature and preliminary studies conducted in December 2009 in 14 participating centers led us to estimate a 1-year mortality after ICU discharge of 18%. To ensure detection with a power of 80% for the detection of binary prognostic factors with a prevalence of 33% and an expected OR of 1.5 in a population with a probability of death in the year following ICU discharge of approximately 18%, 1636 patients were required [ 25]. Assuming a 10% rate of refusal and/or loss to follow-up, the number of patients to be enrolled was raised to 1800. Finally, since the expected in-ICU mortality rate was 25%, the total number of patients included in the study was 2250. p < 0.05 was considered statistically significant. All statistical analyses were performed using R statistical software version 3.1.1 or above (The “R” Foundation for Statistical Computing, Vienna, Austria).

Of the 1570 ICU survivors, 333 (21%) died during the year following ICU discharge, including 123 (8%) during the index hospitalization (Additional file  1: Figure S1). Univariate analysis revealed that the 333 nonsurvivors at 1 year post ICU discharge had a greater degree of illness severity at ICU admission and more comorbidities (Table  1, Additional file  1: Table S1). One-year nonsurvivors were more likely to have septic shock as the cause of admission. While in the ICU, 1-year nonsurvivors required more renal replacement therapy, inotropes/vasopressors and transfusion than survivors. On ICU discharge, nonsurvivors had lower blood pressure and residual organ dysfunction than survivors. Yet renal function was more profoundly altered in nonsurvivors with a higher serum creatinine and lower eGFR at ICU discharge (Table  1).

Multivariable analysis identified 14 independent predictors of post-ICU survival (Fig.  2). Odds ratios of significantly associated variables are presented in Additional file  1: Table S2. Linearity of the association between continuous variables in the multivariable model and the outcome is depicted in Additional file  1: Figure S2. The area of the ROC curve for the multivariable model was 0.787 (95% CI 0.759–0.815). Age and comorbidities (Charlson comorbidity score, vascular disease, severe valvular disease, chronic kidney diseases, cancer and loss of autonomy) were associated with a greater 1-year risk of death. At ICU discharge, five clinical variables (low values of systolic blood pressure, body temperature, total protein and platelet counts, and a high white blood cell count) were associated with an increased post-ICU risk of death. With respect to their ICU stay, red blood cell transfusion and prolonged ICU length of stay were associated with higher risk of 1-year post-ICU mortality. Of note, AUCs of SOFA at admission and SAPS II were 0.574 (95% CI 0.531–0.619) and 0.605 (95% CI 0.572–0.64) respectively; both were significantly lower than the AUC of the clinical score.

At the time of ICU discharge, 1-year nonsurvivors had elevated levels of all measured cardiovascular biomarkers (Table  2). As depicted in Additional file  1: Figure S3, the association between the level of biomarkers at discharge and the outcome was not linear in all cases. After dichotomization according to the median value, elevated biomarkers of cardiac (NT-proBNP, sST2) and vascular (bio-ADM) failure were independently associated with 1-year death when they are added to the multivariable model, with an almost 3-fold increase in the risk of death when combined (adjusted OR 2.84 (95% CI 1.73–4.65), p < 0.001) (Fig.  3). Of note, the association between elevated hs-TnI and 1-year mortality did not remain significant after adjustment. Although only NT-proBNP, bio-ADM and sST2 significantly improve the c-statistic of the clinical model, reclassification analyses showed that all cardiovascular biomarkers, including hs-TnI, improve predictive power of the multivariable model (Table  2).

Sixty-five (4%) patients discharged alive from the ICU were not assessed at 1 year. Although the number is small, this could have affected the accuracy of our results. We cannot assess the risk of readmission after ICU discharge as this information was not recorded prospectively. More broadly, we had no information on patient management (e.g., drug therapy, rehabilitation, psychologist support) after ICU discharge. This may also have contributed to patient vulnerability and needs to be further explored. In addition, while we described clinical and biological variables independently associated with 1-year mortality in ICU survivors, other important parameters need to be considered when discharging a patient from the ICU, such as the amount of nursing care. Some potential predictors of post-ICU outcome were not considered in the present study; in particular, only comorbidities were considered but no frailty score. Moreover, because of the French law, we were not allowed to include patients with no social security coverage, which may limit the external validity of our results. Biological collection was performed when the patient physically left the ICU and not at the time the patient was considered dischargeable from the ICU, which is more tightly linked to the physiologic status of the patient. However, our approach reflects the real-life management of ICU discharge. Although the study was multicentric and conducted in two European countries, only one center outside France included patients; this may limit the external validity of our results. Finally, despite the fact that a sample size calculation was performed, factors that were weakly associated with the 1-year risk of death could not be identified due to insufficient study power. Of note, the main aim of the study was to identify an explanatory model. Thus, the objective of our variable selection procedure was to identify the factors most strongly associated with mortality at 1 year and not to establish a prognostic score that would have to be validated.