The use of maximum SOFA score to quantify organ dysfunction/failure in intensive care. Results of a prospective, multicentre study

Transcription

1 Intensive Care Med (1999) 25: 686±696 Ó Springer-Verlag 1999 ORIGINAL R. Moreno J.-L. Vincent R. Matos A. Mendonça F. Cantraine L. Thijs J. Takala C. Sprung M. Antonelli H. Bruining S. Willatts on behalf of the working group on sepsisrelated problems of the ESICM The use of maximum SOFA score to quantify organ dysfunction/failure in intensive care. Results of a prospective, multicentre study Received: 16 September 1998 Accepted: 16 April 1999 R. Moreno ( ) ) R. Matos Unidade de Cuidados Intensivos Polivalente, Hospital de St. António dos Capuchos, Alameda de St. António dos Capuchos, P-1150 Lisboa, Portugal ( r.moreno@mail.telepac.pt, Fax: + 351± ), J.-L. Vincent A. MendoncËa Department of Intensive Care, Erasme University Hospital, Brussels, Belgium F. Cantraine FacultØ de Medicine, UniversitØ Libre de Bruxelles, Brussels, Belgium L. Thijs Medical Intensive Care Unit, Academisch Ziekenhuis Vrije Universiteit Amsterdam, Amsterdam, The Netherlands J. Takala Department of Intensive Care, Kuopio University Hospital, Kuopio, Finland C. Sprung Department of Intensive Care, Hadassah Hebrew University Medical Center, Jerusalem, Israel M. Antonelli Istituto di Anestesiologia e Rianimazione, Università ªLa Sapienzaº, Rome, Italy H. Bruining Intensive Care Unit, Academisch Ziekenhuis Rotterdam, Rotterdam, The Netherlands S. Willatts Directorate of Anaesthesia, Bristol Royal Infirmary, Bristol, UK Abstract Objective: To evaluate the performance of total maximum sequential organ failure assessment (SOFA) score and a derived measure, delta SOFA (total maximum SOFA score minus admission total SOFA) as a descriptor of multiple organ dysfunction/failure in intensive care. Design: Prospective, multicentre and multinational study. Setting: Forty intensive care units (ICUs) from Australia, Europe, North and South America. Patients: Data on 1,449 patients, evaluated at admission and then consecutively every 24 h until ICU discharge (11,417 records) during May Excluded from data collection were all patients with a length of stay in the ICU less than 2 days following uncomplicated scheduled surgery. Main outcome measure: Survival status at ICU discharge. Interventions: The collection of raw data necessary for the computation of a SOFA score on admission and then every 24 h, and basic demographic and clinical statistics. Measurements and main results: Mean total maximum SOFA score presented a very good correlation to ICU outcome, with mortality rates ranging from 3.2% in patients without organ failure to 91.3 % in patients with failure of all the six organs analysed. A maximum score was reached days after admission for all the organ systems analysed. The total maximum SOFA score presented an area under the ROC curve of (SE 0.012), which was significantly higher than any of its individual components. The cardiovascular score (odds ratio 1.68) was associated with the highest relative contribution to outcome. No independent contribution could be demonstrated for the hepatic score. No significant interactions were found. Principal components analysis demonstrated the existence of a two-factor structure that became clearer when analysis was limited to the presence or absence of organ failure (SOFA score ³ 3 points) during the ICU stay. The first factor comprises respiratory, cardiovascular and neurological systems and the second coagulation, hepatic and renal systems. Delta SOFA also presented a good correlation to outcome. The area under the receiver operating characteristic (ROC) curve was (SE 0.017) for delta SOFA, lower than the total maximum SOFA score or admission total SOFA score. The impact of delta SOFA on prognosis remained significant after correction for admission total SOFA. Conclusions: The results show that total maximum SOFA score and delta SOFA can be used to quantify the degree of dysfunction/failure al-

2 687 ready present on ICU admission, the degree of dysfunction/failure that appears during the ICU stay and the cumulative insult suffered by the patient. These properties make it a good instrument to be used in the evaluation of organ dysfunction/ failure. Key words Severity of illness index Sepsis Multiple organ failure Multiple organ dysfunction syndrome Intensive care Critical care Introduction In recent years, a series of negative results in clinical trials on sepsis [1±13] has challenged the classical adoption of hospital mortality as the end point for the evaluation of clinical trials in intensive care [14]. The use of this measure has constituted the gold standard until now, since it is easy to define and to measure and represents a clinically very relevant end point. However, it has been contested [15], since hospital policy can and does change the location of deaths (e. g. discharging patients to die) and mortality can be significantly underestimated in hospitals which discharge patients very early in the course of their disease. Recently authors have questioned the adequacy of all cause mortality as an end point [16]. A meaningful end point can only be chosen when a direct relation between an event and its consequences is known. In the case of sepsis (and multiple organ failure) our knowledge is very limited and only an indirect and partial relationship can actually be established to most phenomena. Moreover, all cause mortality implies the need for large samples, with problems in reliability of data collection, heterogeneity of enrolled patients and costs. Patients in intensive care, even with strict inclusion criteria for sepsis or septic shock, do not constitute a homogeneous sample. Patients have different diagnoses, time-courses, ages, chronic illnesses (chronic health, co-morbidities), different sites of infection and invading microorganisms and different degrees of physiological dysfunction, resulting in a large dispersion of mortality risks [17, 18]. Additionally, the presence and impact of other confounding events such as inappropriate antimicrobial therapy, inadequate medical-surgical management and forgoing life-sustaining therapies must be analysed and taken into account [19]. Several methods have been proposed to deal with this variation [18, 20, 21], but these usually lead to complex, extensive (and expensive) data collection and sophisticated analysis. It has been shown that certain interventions, effective in their specific scope, fail to reduce all cause hospital mortality. A recent example is selective decontamination of the digestive tract (SDD); it has been demonstrated to diminish the prevalence of nosocomial pneumonia [22], but does not consistently decrease hospital mortality [23, 24]. In other cases, therapeutic interventions have failed to demonstrate beneficial effects in multicentre trials [1] and have been found to be associated with a significant reduction in mortality when applied to more homogeneous groups of patients [25]. The awareness of these factors led the Working Group on Sepsis-related Problems of the European Society of Intensive Care Medicine (ESICM) to organise a consensus meeting in Paris (December 1994) to create the so-called sepsis-related organ failure assessment (SOFA) score [26], later called sequential organ failure score since it is not restricted to sepsis. The rationale behind this decision was the necessity to find an objective and simple way to describe individual organ dysfunction/failure in a continuous form, from mild dysfunction to severe failure, that can be used over time to measure the evolution of individual (or aggregated) organ dysfunction in clinical trials on sepsis or for the clinician at the bedside. A retrospective evaluation of the application of this score to the first 24 h in the ICU on 1,643 patients with early sepsis on an international database [26] demonstrated a good correlation to mortality and an acceptable distribution of the patients among the several groups. To confirm these retrospective findings, a prospective, multinational study was initiated, the main results of which are presented elsewhere [27]. The aim of this work was to evaluate the performance of total maximum SOFA score and a derived measure, delta SOFA (total maximum SOFA minus admission total SOFA, that is, the magnitude of organ dysfunction appearing during the ICU stay) as a descriptor of multiple organ dysfunction/failure in intensive care. Materials and methods Patients Two months before the start of data collection, all participants in the working group on sepsis-related problems of the ESICM were invited to collaborate. Data collection took place from May 1, 1995, to May 31, Forty ICUs from 16 countries in Australia (1), Europe (35), North (1) and South America (3) participated in the study. Patients with a length of stay (LOS) in the ICU less than 2 days following uncomplicated scheduled surgery were excluded from data collection. For each patient a simple set of variables was collected that included basic demographic characteristics and all the variables of the SOFA score [26] (see Appendix). Data were registered on admission and every 24 h thereafter using the worst values until ICU discharge. All data were collected as raw data. In the co-ordinating centre (Erasme Hospital, Free University of Brussels, Belgium), data were entered into a computer format using a continuous pro-

3 688 Fig. 1 Frequency distribution of maximum SOFA in survivors (open bars, n = 1131) and non-survivors (black bars, n = 313). Asterisks present the relationship between maximum SOFA score and ICU mortality, with the logistic regression curve superimposed SOFA score to prognosis, using outcome in the ICU as the dependent variable and delta SOFA and admission SOFA as the independent variables. Logistic regression analysis was used to evaluate the relationship between total maximum SOFA score and ICU mortality. Linear regression analysis was used to evaluate the correlation between mean delta SOFA and ICU mortality. Exploratory factor analysis was applied to study meaningful interrelations among the six components of the SOFA score. This type of analysis aims at analysing the interrelations between a set of variables, without the need to consider which variables are dependent and which variables are independent (opposite to regression where this specification must be done). This technique allows for the identification of association patterns of the different variables under consideration, without first having to specify a causeand-effect relationship [30]. An eigen value of ³ 1 or more was considered as significant. The results are presented as means standard deviation except when stated otherwise. The outcome measure used was survival status at discharge from the ICU. Data analysis and statistics were performed using the Statistical Package for Social Sciences (SPSS) version 5.0 for MS DOS and 7.0 for Microsoft Windows at the Intensive Care Unit, Hospital de Santo António dos Capuchos, Lisbon, Portugal. cess of monitoring of its completeness and correction. For a single missing value a replacement was calculated using the mean value of the result preceding and that following the missing one. When more than one consecutive value was missing it was considered as a missing value in the analysis. More details about data collection are given elsewhere [27]. Methods Maximum organ failure scores were calculated for all the six components of the system during the entire ICU stay. The aggregate score (total maximum SOFA score) was calculated summing the worst scores for each of the components. The amount of organ dysfunction/failure appearing after ICU admission (delta SOFA) was evaluated computing the total maximum SOFA score minus the admission total SOFA score (Appendix). For purposes of analysis, organ dysfunction was defined as a SOFA score of 1 or 2 points and organ failure as a SOFA score ³ 3. Chi-square statistics (with Yates correction when applicable) were used to test for the statistical significance of categorical variables and one-way analysis of variance was used to assess continuous variables. All statistical tests were two-sided, and a significance level of 0.05 or less was used except when stated otherwise. The discriminative power of the scores, that is, the ability of the scores to discriminate between patients who live and patients who die, was defined by the area under the receiver operating characteristic (ROC) curve, computed by a modification of the Wilcoxon statistics, as proposed by Hanley and McNeil [28]. The comparison of the areas under ROC curves was made using the Z statistic with correction for the correlation introduced by studying the same sample [29]. For the computation of the odds ratios and interactions associated with each component of the system, we fitted a logistic regression model with outcome in the ICU as the dependent variable. The maximum SOFA scores for each of the six systems were used as independent variables. Later, pertinent interactions were added to the model. The same method was applied in the evaluation of the relative contributions of delta SOFA and admission total Results A total of 1,449 patients were studied for a total of 11,417 ICU days. Thirty-two percent of the patients were admitted from the emergency room, 27% from the ward, 26 % from the operative theatre and 11% from other hospitals. Most patients were male (64 %), with an overall mean age of years ranging from 12 to 95 years. Non-operative patients comprised 44% of the sample. The median LOS in the ICU was 5.0 days (interquartile range 3±10 days). The overall mortality in the ICU was 22% with a corresponding overall hospital mortality of 26 %. Five patients had missing data in their ICU outcome and were excluded from the analysis related to outcome. More details can be found in the main description of the study [27]. Figure 1 shows the frequency distribution of total maximum SOFA score in survivors and non-survivors. The mean total maximum SOFA score was points, median 7 points, range 0±24 points, and was significantly higher in non- survivors than in survivors ( points versus points, p < 0.001). The relationship between total maximum SOFA score and ICU outcome was established as: Pr = e 4:0473+0:2790(TMS) 1+e 4:0473+0:2790(TMS) where Pr is the probability of death in the ICU and TMS is the total maximum SOFA score during the ICU stay. In all the organ systems analysed, the maximum SOFA score during the ICU stay for that specific organ presented an acceptable frequency distribution among the various groups (Fig. 2). The differences between sur-

4 689 Table 1 Maximum SOFA scores for the six organ systems in the global population, in survivors and in non-survivors. The differences between survivors and non-survivors were always significant (p < 0.001). Results are presented as mean ± standard deviation Component Global population (n = 1,444) Survivors (n = 1,131) Non-survivors (n = 313) Respiratory 2.2 ± ± ± 1.1 Cardiovascular 1.5 ± ± ± 1.4 Renal 1.0 ± ± ± 1.4 Coagulation 1.0 ± ± ± 1.3 Hepatic 0.7 ± ± ± 1.3 Neurological 1.7 ± ± ± 1.6 Total Maximum SOFA score 8.2 ± ± ± 4.8 Table 2 Number of organ failures (maximum SOFA score L 3 points) and ICU outcome Number of organ failures Number of patients %of patients Death in the ICU (n) Mortality rate (%) Maximum SOFA score a ± ± ± ± ± ± ± 1.5 a mean ± standard deviation Fig. 2 Mortality rate (bars) and number of patients in each organ system (*), according to maximum SOFA score vivors and non-survivors were always significant (Table 1). The mortality rate increased as the score increased in all organ systems (Fig. 2). For respiratory and neurological scores the patterns were less clear, especially when the scores were lower than 3 points. The number of organ failures (maximum SOFA score ³ 3) also showed a significant correlation to ICU outcome, with mortality rates ranging from 3.2 % in patients without any organ failure to 91.3 % in patients with failure of all the six organs analysed (Table 2). The mean maximum SOFA score also showed a corresponding increase, with values ranging from points in patients without organ failures to points in patients with six organ failures. The maximum SOFA score occurred shortly after admission for all organ systems analysed ( days) with mean values ranging from 0.8 days (95 % confidence interval 0.6±0.9 days) for the neurological score to 1.4 days (95 % confidence interval 1.2±1.5 days) for the respiratory score. If we limit the analysis to organ failures (SOFA ³ 3 points), the time required to reach maximum values was longer (mean days, ranging from 1.6 days for the neurological score to 4.9 for liver failure) than the time needed to reach a maximum SOFA score, and significant differences were noted between the different organ systems analysed (p < 0.001). The neurological system was the first to fail, the respiratory, cardiovascular, renal and coagulation systems occupied an intermediate position and the hepatic was the last (Fig. 3). In the evaluation of the discriminative power of the scores, the area under the ROC curve was used. The best discriminative power was shown for the cardiovascular score (0.802, standard error (SE) 0.015), the renal score (0.739, SE 0.016) and the respiratory score (0.736, SE 0.016). For the neurological score the value was intermediate (0.727, SE 0.016). Coagulation (0.684, SE 0.018) and hepatic scores (0.655, SE 0.019) had a lower discriminative power. The aggregated score (total maximum SOFA score) presented an area under the ROC curve of (SE 0.012) which was significantly higher (cardiovascular score p = 0.005, all others p < 0.001) than any of its individual components (Fig. 4). In order to evaluate the relative contribution to outcome of each of the six individual organ system dysfunctions, a non-stepwise logistic regression equation was developed, relating the score for each organ to the outcome in the ICU. In this way, the exponent of the esti-

5 690 Fig.3 Mean time to reach maximum SOFA score in patients with organ failure (SOFA ³ 3). Values are presented as mean 95% confidence intervals for the mean. The numbers of patients in each of the systems were: respiratory 644, cardiovascular 392, renal 198, coagulation 190, hepatic 119 and neurological 561 Fig. 4 Discriminative power of total maximum SOFA score, delta SOFA and admission total SOFA score. The receiver operating characteristics (ROC) curve summarises the relationship between sensitivity (number of true positives) and 1 minus specificity (number of false positives) for all the possible values of the score. The reference line represents the discriminative power of a score no better than chance (area under ROC curve 0.5). The values within brackets are standard errors. The outcome used was vital status at ICU discharge mated coefficient (b) for each organ score represents the factor by which the odds ratio of ICU death changes when the score for that particular organ increases 1 point. The results (Table 3) demonstrated that the cardiovascular score was associated with the highest relative contribution to outcome (odds ratio 1.68, 95% confidence interval 1.49±1.91), followed by the renal (odds ratio 1.46, 95 % confidence interval 1.29±1.64), the neurological (odds ratio 1.40, 95 % confidence interval 1.28±1.55), the coagulation (odds ratio 1.22, 95% confidence interval 1.06±1.40) and the respiratory (odds ratio 1.18, 95 % confidence interval 1.01±1.38) scores. No such contribution could be demonstrated for the hepatic score (odds ratio 0.82, 95 % confidence interval 0.60±1.11). The same technique was used to evaluate significant interactions among the six components of the SOFA score. However, no conclusive results could be demonstrated. A trend (non-significant) was found for interactions between respiratory and coagulation scores (odds ratio 1.14, 95% confidence interval 0.97±1.33), respiratory and renal scores (odds ratio 1.09, 95% confidence interval 0.95±1.25), cardiovascular and renal scores (odds ratio 1.07, 95% confidence interval 0.96±1.07), renal and hepatic scores (odds ratio 1.11, 95% confidence interval 0.98±1.25), and hepatic and coagulation scores (odds ratio 1.09, 95% confidence interval 0.97±1.23). Exploratory factor analysis was then used to identify meaningful interrelations among the six components Table 3 Relative contributions to ICU outcome of the maximum value during ICU stay for each of the six components of the SOFA score Variable b SE Wald p R Odds-ratio (95% confidence intervals) Respiratory (1.007±1.378) Cardiovascular < (1.488±1.905) Renal < (1.294±1.643) Coagulation (1.059±1.404) Hepatic Ÿ (0.603±1.107) Neurological < (1.275±1.545) Constant Ÿ b, coefficient; SE, standard error; Wald, Wald statistic, R, partial correlation. Odds-ratios are presented for a 1-point change in the scores for each organ

6 691 Table 4 Principal components analysis of the six components of the SOFA system. Results are presented for maximum values during ICU stay (top) and for the presence of organ failure (SO- FA L 3 points) during ICU stay (bottom), after varimax rotation Component Factor 1 Factor 2 Maximum values during ICU stay Respiratory Cardiovascular Renal Coagulation Hepatic Neurological Organ failures (SOFA L 3 points) during ICU stay Respiratory Cardiovascular Renal Coagulation Hepatic Neurological Fig. 5 Delta SOFA score and ICU mortality. A linear relation exists between delta SOFA and ICU mortality Table 5 Delta SOFA (total maximum SOFA score minus admission total SOFA score) and ICU outcome Delta SOFA Number of patients %of patients Death in the ICU (n) Mortality rate (%) L of the SOFA score. The results (Table 4) demonstrated the existence of a two-factor structure, each comprising three components of the system, that became more clear when the analysis was limited to the presence or absence of organ failure (SOFA score ³ 3) during the ICU stay. The first factor comprises the respiratory, cardiovascular and neurological scores and the second the coagulation, hepatic and renal components. Delta SOFA presented a mean value of points, median 2.0 points, range 0±19 points. Delta SOFA was significantly higher (p < 0.001) in non-survivors than in survivors ( points versus points). As presented in Table 5, ICU mortality increased as the delta SOFA score increased. The association between mean delta SOFA and ICU mortality followed a linear pattern (Fig. 5, Table 5). Delta SOFA presented an area under the ROC curve of (SE 0.017) (Fig. 4), which was significantly lower (p < 0.001) than that of the total maximum SOFA score and slightly lower (non-significant) than that of the admission total SOFA score (0.772, SE 0.015). In order to evaluate the relative contribution to ICU outcome of the amount of organ dysfunction present at ICU admission (admission total SOFA score) and that developing during ICU stay (delta SOFA score), a nonstepwise logistic regression equation was developed. Results demonstrate that both were important for outcome, and with a similar weight (Table 6). The associated odds ratios, for a 1 point change in the score were 1.36 (95 % confidence interval 1.30±1.42) for the admis- Table 6 Relative contribution for outcome in the ICU of the admission SOFA score and delta SOFA Variable b SE Wald p R Odds-ratio (95% confidence intervals) Admission total SOFA < (1.303±1.421) score Delta SOFA < (1.303±1.432) Constant Ÿ b, coefficient; SE, standard error; Wald, Wald statistic, R, partial correlation. Odds-ratios are presented for a 1-point change in the score.

7 692 sion total SOFA score and 1.37 (95 % confidence interval 1.30±1.43) for the delta SOFA score. Discussion Multiple organ dysfunction syndrome (MODS) has become the leading cause of morbidity and mortality in intensive care [31±33]. Described initially by Tilney et al. in 1973 after massive acute blood loss and shock [34], it was found later to be associated with infection [35, 36], acute pancreatitis [37], burns [38], shock [39] and trauma [40]. As emphasised by a recent Consensus Conference [41], there is a need for a comprehensive database to test and validate optimal criteria for describing this syndrome, in which specific variables could be tested against outcome. Various efforts to this end have appeared recently in the literature [26, 42±44]. All were built on the common assumptions: that one can describe increasing dysfunction in individual organs and assess MODS as a continuum of organ dysfunction/failure instead of an on/off phenomenon. However, limitations exist for all. The systems proposed by Marshall et al. (multiple organ dysfunction score) and by Bernard et al.(brussels score) have not been tested in a multicentre representative database of critically ill patients. The system proposed by Le Gall et al., logistic organ dysfunction (LOD) score, was developed with very sophisticated statistical techniques to choose and weigh the variables in a large international database. However, it was developed and validated with data collected only in the first 24 h in the ICU and no information exists about its behaviour at later stages in the evolution of MODS. A panel of experts constructed the latest system, SOFA score, based on a review of the literature. This methodology, has been applied successfully in the past [45] but needs extensive validation in order to evaluate the adequacy of the variables chosen and their limits. This was recognised in the original description [26] and prompted the Working Group on Sepsis-related Problems of the ESICM to perform a prospective, multinational validation study. The main data have been presented elsewhere [27]. Based on these data, we studied the validity of two complementary measures as descriptors of morbidity in intensive care: total maximum SOFA score and delta SOFA. The results show that, in this ICU patient database, total maximum SOFA score showed a very good correlation to outcome and occurred early during the ICU stay. All the individual organ scores were significantly higher in non-survivors than in survivors, with a clear correlation between increasing score and increasing mortality except for low values (less than 3) of the neurological and respiration scores, where the patterns were not clear. The same relation was present when we limited the analysis to organ failure (SOFA score ³ 3 points). The discriminative power was very good (area under ROC curve 0.847, SE 0.012). For individual organ scores, the best discriminative power was seen for cardiovascular score. In multivariate analysis the impact on outcome of organ dysfunction/failure was higher for cardiovascular (odds ratio 1.68) and renal (odds ratio 1.46) scores. The hepatic dysfunction/failure did not show a significant impact on prognosis (odds ratio 0.82). No significant interactions were seen between individual organ failures. Using principal components analysis, a clear pattern was seen when we analysed organ failures (SOFA score ³ 3 points), with a two-factor structure: respiratory, cardiovascular and neurological, and coagulation, hepatic and renal. In the overall analysis the same pattern was present although less clear, with the cardiovascular score having an intermediate position. The amount of organ dysfunction/failure occurring after ICU admission (delta SOFA) also showed a good correlation to outcome. On multivariate analysis this effect was still significant after controlling for admission score. It should be noted that the delta SOFA ability to distinguish between patients who died and patients who survived was lower than that of the total maximum SOFA score or even than of the admission SOFA score. This stresses the importance of the degree of physiological derangement on admission to the ICU [46±48] and of cumulative organ dysfunction [31, 43] to the prognosis. Why use a total maximum SOFA score instead of a simpler measure? Our rationale was that a daily evaluation would not be able to capture the overall amount of organ dysfunction/failure sustained by the patient during the course of the disease. Different organs are affected in this complex physiopathological process at different points in time [32] and a daily evaluation, although appealing, can miss the total amount of organ dysfunction sustained by the patient, leading to an underestimation of the cumulative insult suffered. It has been shown that mortality due to MODS depends on the number of failing organs [31, 43, 49], on the severity of the dysfunction/failure [43, 44], on the particular combination of failing organs [49±51] and on the duration [31, 49]. Our system, following the path of previous work by Marshall et al. [43], allows the quantification of all these conditions. Alternative approaches, based on the daily application of severity scores have been proposed [48, 52±57] but are usually limited to the first days in the ICU [48, 54] or have later failed to confirm their initial performances [58]. Additionally, the proposed system allows the distinction between the dysfunction/failure already present at ICU admission (which depends mainly on admission policies), the dysfunction/failure that appears during

8 693 the ICU stay and the evaluation of the total insult suffered by the patient. All are very important by themselves, as shown in Table 6, but also address complementary facets of a complex response. The admission SOFA reflects the degree of failure already present when the patient enters the ICU. This measurement, that only the admission mortality prediction model [47] is able to achieve, can be used to stratify patients according to severity of illness, for example, for inclusion in clinical trials based on the admission SOFA score. The delta SOFA measures the progress of the patient during the ICU stay and is potentially influenced by therapy. The fact that it was a good prognostic indicator after controlling for admission SOFA score suggests that strategies directed at the prevention and/or limitation of further organ dysfunction will have a significant impact on prognosis, independent of the condition of the patient on admission to the ICU. This certainly needs further research. Last but not least, the quantification of the total insult suffered by the patient during the ICU stay (total maximum SOFA) was a very important prognostic indicator. This suggests that it can be used to quantify the impact of therapeutic interventions on overall or organ-specific morbidity. Some but not all of those interventions could also have an impact on mortality, but to focus exclusively on mortality as an end point could lead to an underestimation of the relevant effects of therapeutic interventions obscured by the heterogeneity of causes of death. What is the precise nature of the two-factor structure observed? What are the precise relationships between the respiratory, cardiovascular and neurological systems or between the coagulation, hepatic and renal systems? One tempting explanation could be the presence of two targets in this complex syndrome. If this is the case, the first association would represent the primary insult (e. g. shock or severe respiratory failure) and the second its late consequences, appearing as a result of the host response to the primary insult. This two-target explanation is consistent with previous descriptions [41, 59] but must be tested in adequate models. The presence of neurological dysfunction/failure in the first factor could be explained by the presence of patients with trauma in this database (181 patients, although probably not all with head trauma) or by the early onset of septic encephalopathy in MODS [60]. Moreover, concerns about the reliability of the evaluation of neurological dysfunction in critically ill patients have recently been raised [61], although not shared by all the researchers [62, 63]. Maybe when physiologists return from the drawing board, as recently suggested [64], we will gain more insight into the explanation of this phenomenon. Our study presents some limitations that must be acknowledged. First, we only evaluated the relationship of SOFA with ICU outcome and not with hospital or 30-day mortality. This fact could have introduced some bias in the analyses and more research should be undertaken to examine whether there exists a link between organ dysfunction/failure during the ICU stay, shortterm (ICU) mortality and long-term mortality. For that purpose, patients must be followed after ICU discharge and monitored for the development of further complications. Second, SOFA, similar to all the published organ failure scores, uses the Glasgow coma score for neurological evaluation [65] and this computation can be very difficult or impossible in sedated patients and very prone to errors in data collection. Certainly we need to develop better ways to assess neurological dysfunction in the critically ill, non-trauma patient. The best treatment for MODS is certainly prevention. Unfortunately, this is not possible in many cases. New diagnostic tools and new therapeutic options are needed to deal with this complex syndrome that is responsible for so many deaths. In the meantime, instruments like the SOFA score and their derived measures should be used for the evaluation and quantification of organ dysfunction/failure. Acknowledgements The authors want to acknowledge the efforts in data collection by all the participants in the study. A complete list of participating centres can be found in J.-L. Vincent et al. [27]. Appendix SOFA score was computed at admission and for every 24 h period from the most deranged values for each of the organ systems considered [26]. An example of the computation of the associated values is shown below: A patient was admitted to the ICU subsequent to surgery for a perforated duodenal ulcer, complicated by peritonitis. At admission, he had respiratory failure with a PaO 2 /FiO 2 ratio of 180 on mechanical ventilation, mild cardiovascular dysfunction (mean arterial pressure 60 mmhg without vasoactive drugs), and mild neurological dysfunction (Glasgow Coma score 14). There were no renal, liver or coagulation disturbances (blood creatinine 1.0 mg/dl, serum bilirubin 1.0 mg/dl and platelets/mm 3 ). The SOFA score computed at admission was 5 points. During his ICU stay, the respiratory function improved with the patient being weaned from the ventilator on day 2 and presenting a PaO 2 /FiO 2 ratio of 420 on the day of discharge. Cardiovascular support with dobutamine was needed on days 1 and 2. A mild renal dysfunction (creatinine 1.6) was present on days 1 and 2. Thrombocytopenia (minimal value platelets/ mm 3 ) and hyperbilirubinaemia (maximum serum bilirubin 7.8 mg/dl) appeared during the ICU stay. Neurological function worsened during days 2 and 3 (Glasgow

9 694 SOFA score Respiration PaO 2 /FiO 2 mmhg < 400 < 300 < 200 with respiratory support Coagulation < 150 < 100 < 50 < 20 Platelets x 10 3 /mm 3 Liver Bilirubin, mg/dl (mmol/l) 1.2±1.9 (20±32) 2.0±5.9 (33±101) Cardiovascular MAP < 70 mm Hg Dopamine K 5or Hypotension a Dobutamine (any dose) Central Nervous System Glasgow coma score Renal Creatinine, mg/dl (mmol/l) or urine output 6.0±11.9 (102±204) Dopamine < 5 or epinephrine K 0.1 or norepinephrine K ±14 10±12 6±9 < 6 1.2±1.9 (110±170) 2.0±3.4 (171±299) a adrenergic agents administered for at least one hour (doses given are in mg/kg min) 3.5±4.9 (300±440) or < 500 ml/day < 100 with respiratory support > 12.0 (> 204) Dopamine > 1.5 or epinephrine > 0.1 or norepinephrine > 0.1 > 5.0 (> 440) or < 200 ml/day coma score 12) and then improved, with a Glasgow Coma Score of 15 at discharge. The patient was discharged to the ward on day 5, still with thrombocytopenia and hyperbilirubinaemia. The summary of the evolution of the patient in terms of SOFA score is given below. Total maximum SOFA score was 14 points, and delta SOFA score 9 points The summary of the evolution of the patient in terms of SOFA score is given bellow Day 0 Day 1 Day 2 Day 3 Day 4 Day 5 Respiratory Cardiovascular Renal Coagulation Hepatic Neurological Total maximum SOFA score was 14 points, and delta SOFA score 9 points. References 1. Bone RC, Fisher CJ, Clemmer TP, et al. (1987) A controlled trial of high-dose methylprednisolone in the treatment of severe sepsis and septic shock. N Engl J Med 317: 653± Greenman RL, Schein RMH, Martin MA, et al. (1991) A controlled clinical trial of E5 murine monoclonal IgM antibody to endotoxin in the treatment of Gram-negative sepsis. JAMA 266: 1097± Ziegler EJ, Fisher CJ, Sprung CL, et al. (1991) Treatment of gram-negative bacteremia and septic shock with ha-1 a human monoclonal antibody against endotoxin. A randomized, double-blind, placebo-controlled trial. N Engl J Med 324: 429± Dhainaut J-FD, Tenaillon A, Le Tulzo Y, et al. (1994) Platelet-activating factor receptor antagonist BN in the treatment of severe sepsis: a randomized, double-blind, placebo-controlled, multicenter clinical trial. Crit Care Med 22: 1720± Fisher CJ, Dhainaut J-FD, Opal SM, et al. (1994) Recombinant human interleukin-1 receptor antagonist in the treatment of patients with sepsis syndrome: results from a randomized, double-blind, placebo-controlled trial. JAMA 271: 1836± Abraham E, Wunderink R, Silverman H, et al. (1995) Efficacy and safety of monoclonal antibodies to human necrosis factor-alpha in patients with sepsis syndrome. JAMA 273: 934± Cohen J, Carlet J (1996) INTERSEPT: an international, multicenter, placebocontrolled trial of monoclonal antibody to human tumor necrosis factor-alpha in patients with sepsis. International Sepsis Trial Study Group. Crit Care Med 24: 1431± Fisher CJ Jr, Agosti JM, Opal SM, et al. (1996) Treatment of septic shock with the tumor necrosis factor receptor:fc fusion protein. N Engl J Med 334: 1697± Froom AM, Greve JW, Buurman WA, et al. (1996) Treatment with the platelet-activating factor antagonist TCV- 309 in patients with severe systemic inflammatory response syndrome: a prospective, multi-center, double-blind, randomized phase II trial. Shock 5: 313± Reinhart K, Wiegand-Loehnert C, Grimminger F, et al. (1996) Assessment of the safety and efficacy of the monoclonal anti-tumour necrosis factor antibody-fragment MAK 195F, in patients with sepsis and septic shock: a multicenter, randomized, placebo-controlled, dose-ranging study. Crit Care Med 24: 733± Abraham E, Glauser MP, Butler T, et al. (1997) p55 Tumor necrosis factor fusion protein in the treatment of patients with severe sepsis and septic shock. A randomized controlled multicenter trial. JAMA 277: 1531±1538

10 Bernard GR, Wheeler AP, Russell JA, et al. (1997) The effects of ibuprofen on the physiology and survival of patients with sepsis. N Engl J Med 336: 912± Fein AM, Bernard GR, Criner GJ, et al. (1997) Treatment of severe systemic inflammatory response syndrome and sepsis with a novel bradykinin antagonist, deltibant (CP-0127). Results of a randomized, double-blind, placebocontrolled trial. JAMA 277: 482± Sibbald WJ, Vincent J-L (1995) Round table conference on clinical trials for the treatment of sepsis. Brussels, March 12±14, Intensive Care Med 21: 184± Jencks SF, Williams DK, Kay TL (1988) Assessing hospital-associated deaths from discharge data. The role of length of stay and comorbidities. JAMA 260: 2240± Petros AJ, Marshall JC, Van-Saene HK (1995) Should morbidity replace mortality as an endpoint for clinical trials in intensive care? Lancet 345: 369± Knaus WA, Wagner DP, Harrell FE, Draper EA (1994) What determines prognosis in sepsis? Evidence for a comprehensive individual patient risk assessment approach to the design and analysis of clinical trials. In: Reinhart K, Eyrich K, Sprung C (eds) Sepsis. Current Perspectives in Pathophysiology and Therapy. Springer, Berlin Heidelberg, pp 23± Le Gall J-R, Lemeshow S, Leleu G, et al. (1995) Customized probability models for early severe sepsis in adult intensive care patients. JAMA 273: 644± Sprung CL, Finch RG, Thijs LG, Glauser MP (1996) International sepsis trial (INTERSEPT): role and impact of a clinical evaluation committee. Crit Care Med 24: 1441± Knaus WA, Harrell FE, Fisher CJ, et al. (1993) The clinical evaluation of new drugs for sepsis. A prospective study design based on survival analysis. JAMA 270: 1233± Knaus WA (1994) Principles of severity stratification and outcome prediction in sepsis and shock. Intensive Care Med 20: S Stoutenbeek CP, Van Saene HK, Miranda DR, Zandstra DF (1984) The effect of selective decontamination of the digestive tract on colonisation and infection rate in multiple trauma patients. Intensive Care Med 10: 185± Kollef MH (1994) The role of selective digestive tract decontamination on mortality and respiratory tract infections. A meta-analysis. Chest 105: 1101± GrØgoire G, Derderian F, Le Lorier J (1995) Selecting the language of the publications included in a meta-analysis: is there a tower of Babel bias? J Clin Epidemiol 48: 159± Hoffman SL, Punjabi NH, Kumala S, et al. (1984) Reduction of mortality in chloramphenicol treated severe thyphoid fever by high-dose dexamethasone. N Engl J Med 310: 82± Vincent J-L, Moreno R, Takala J, et al. (1996) The SOFA (Sepsis-related organ failure assessment) score to describe organ dysfunction/failure. Intensive Care Med 22: 707± Vincent J-L, De Mendonça A, Cantraine F, et al. (1998) Use of the SOFA score to assess the incidence of organ dysfunction/failure in intensive care units: results of a multicentric, prospective study. Crit Care Med 26: 1793± Hanley J, McNeil B (1982) The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 143: 29± Hanley J, McNeil B (1983) A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology 148: 839± Kleinbaum DG, Kupper LL, Muller KE (1987) Applied regression analysis and other multivariate methods. PWS-Kent Publishing Company, Boston 31. Knaus WA, Draper EA, Wagner DP, Zimmerman JE (1985) Prognosis in acute organ-system failure. Ann Surg 202: 685± Deitch EA (1992) Multiple organ failure: pathophysiology and potential future therapy. Ann Surg 216: 117± Tran DD (1994) Age, chronic disease, sepsis, organ system failure and mortality in the elderly admitted to the intensive care medicine. Intensive Care Med 20: S Tilney NL, Baily GL, Morgan AP (1973) Sequential system failure after rupture of abdominal aortic aneurysms: an unsolved problem in postoperative care. Ann Surg 178: 117± Fry DE, Pearlstein L, Fulton RL, Polk HC (1980) Multiple system organ failure. The role of uncontrolled infection. Arch Surg 115: 136± Bell RC, Coalson JJ, Smith JD, Johanson WG (1983) Multiple organ failure and infection in adult respiratory distress syndrome. Ann Intern Med 99: 293± Tran DD, Cuesta MA (1992) Evaluation of severity in patients with acute pancreatitis. Am J Gastroenterol 87: 604± Marshall WG, Dimick AR (1983) Natural history of major burns with multiple subsystem failure. J Trauma 23: 102± Henao FJ, Daes JE, Dennis RJ (1991) Risk factors for multiorgan failure: a case-control study. J Trauma 31: 74± Faist E, Baue AE, Dittmer H, Heberer G (1983) Multiple organ failure in polytrauma patients. J Trauma 23: 775± Members of the American College of Chest Physicians and the Society of Critical Care Medicine Consensus Conference Committee (1992) American College of Chest Physicians / Society of Critical Care Medicine Consensus Conference: definitions of sepsis and multiple organ failure and guidelines for the use of innovative therapies in sepsis. Crit Care Med 20: 864± Bernard GR, Doig BG, Hudson G, et al. (1995) Quantification of organ failure for clinical trials and clinical practice. Am J Respir Crit Care Med 151: A Marshall JC, Cook DA, Christou NV, Bernard GR, Sprung CL, Sibbald WJ (1995) Multiple organ dysfunction score: a reliable descriptor of a complex clinical outcome. Crit Care Med 23: 1638± Le Gall JR, Klar J, Lemeshow S, et al. (1996) The logistic organ dysfunction system. A new way to assess organ dysfunction in the intensive care unit. JAMA 276: 802± Knaus WA, Draper EA, Wagner DP, Zimmerman JE (1985) APACHE II: a severity of disease classification system. Crit Care Med 13: 818± Knaus WA, Wagner DP, Draper EA, et al. (1991) The APACHE III prognostic system. Risk prediction of hospital mortality for critically ill hospitalized adults. Chest 100: 1619± Lemeshow S, Teres D, Klar J, Avrunin JS, Gehlbach SH, Rapoport J (1993) Mortality Probability Models (MPM II) based on an international cohort of intensive care unit patients. JAMA 270: 2478± Wagner DP, Knaus WA, Harrel FE Jr, Zimmerman JE, Watts C (1994) Daily prognostic estimates for critically ill adults in intensive care units: results from a prospective, multicenter, inception cohort analysis. Crit Care Med 22: 1359± Zimmerman JE, Knaus WA, Sun X, Wagner DP (1996) Severity stratification and outcome prediction for multisystem organ failure and dysfunction. World J Surg 20: 401±405

11 Hebert PC, Drummond AJ, Singer J, Bernard GR, Russell JA (1993) A simple multiple system organ failure scoring system predicts mortality of patients who have sepsis syndrome. Chest 104: 230± Fagon J-Y, Chastre J, Novara A, Medioni P, Gilbert C (1993) Characterization of intensive care unit patients using a model based on the presence or absence of organ dysfunctions and/or infection: the ODIN model. Intensive Care Med 19: 137± Chang RWS, Jacobs S, Lee B (1988) Predicting outcome among intensive care unit patients using computerised trend analysis of daily Apache II scores corrected for organ system failure. Intensive Care Med 14: 558± Chang RWS (1989) Individual outcome prediction models for intensive care units. Lancet i: 143± Lemeshow S, Klar J, Teres D, et al. (1994) Mortality probability models for patients in the intensive care unit for 48 or 72 hours: a prospective, multicenter study. Crit Care Med 22: 1351± Rogers J, Fuller HD (1994) Use of daily acute physiology and chronic health evaluation (APACHE) II scores to predict individual patient survival rate. Crit Care Med 22: 1402± Yzerman EP, Boelens HA, Tjhie JH, Kluytmans JA, Mouton JW, Verbrugh HA (1996) Delta APACHE II for predicting course and outcome of nosocomial Staphylococcus aureus bacteremia and its relation to host defense. J Infect Dis 173: 914± Pittet D, ThiØvent B, Wenzel RP, Li N, Auckenthaler R, Suter PM (1996) Bedside prediction of mortality from bacteriemic sepsis. A dynamic analysis of ICU patients. Am J Respir Crit Care Med 153: 684± Jacobs S, Arnold A, Clyburn PA, Willis BA (1992) The Riyadh intensive care program applied to a mortality analysis of a teaching hospital intensive care unit. Anaesth Resusc Intensive Ther 47: 775± Redl H, Schlag G (1990) Pathophysiology of multi-organ failure (MOF) ± proposed mechanisms. Clin Intensive Care 1: 66± Eidelman LA, Putterman D, Putterman C, Sprung CL (1996) The spectrum of septic encephalopathy. Definitions, etiologies, and mortalities. JAMA 275: 470± Rowan K (1996) The reliability of case mix measurements in intensive care. Curr Opin Crit Care 2: 209± Damiano AM, Bergner M, Draper EA, Knaus WA, Wagner DP (1992) Reliability of a measure of severity of illness: acute physiology and chronic health evaluation -II. J Clin Epidemiol 45: 93± Bastos PG, Sun X, Wagner DP, Wu AW, Knaus WA (1993) Glasgow coma scale score in the evaluation of outcome in the intensive care unit: findings from the acute physiology and chronic health evaluation III study. Crit Care Med 21: 1459± Chernow B (1996) Back to the drawing board. Crit Care Med 24: 1097± Bernard GR (1998) Quantification of organ dysfunction: seeking standardization. Crit Care Med 26: 1767±1768