| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Eur J Cardiothorac Surg 2003;24:741-749
© 2003 Elsevier Science NL
a Department of Medical Informatics, Academic Medical Center, Room J2-263, P.O. Box 22700, 1100 DE Amsterdam, The Netherlands
b Department of Anesthesia and Intensive Care, Amphia Hospital, Breda, The Netherlands
c Department of Cardiology, Academic Medical Center, 1100 DE Amsterdam, The Netherlands
Received 16 April 2003; received in revised form 19 July 2003; accepted 21 July 2003.
* Corresponding author. Tel.: +31-20-566-6893; fax: +31-20-691-9840
e-mail: r.v.huijskes{at}planet.nl
 |
Abstract |
|---|
 Top Abstract 1. Introduction2. Material and methods3. Results4. DiscussionReferences |
|---|
Key Words: Outcome prediction • In-hospital death • Major adverse cardiac events • Extended length of stay intensive care • Euroscore
|
1. Introduction |
|---|
|
TopAbstract1. Introduction2. Material and methods3. Results4. DiscussionReferences |
|---|
When making the decision to perform cardiac surgery, the physician assesses the patient's risk for death and complications. Perioperative death and complication rates, after adjustment for preoperative patient characteristics, are also used for benchmarking between hospitals and quality assurance. Several risk stratification models for cardiothoracic surgery have been developed, like the Parsonnet model [1] and, more recently, the Euroscore [2]. The Parsonnet model, developed in the late 80s in the United States, tends to overpredict perioperative death [3–5]. The Euroscore is a more recently developed model that has been validated in different centres all over Europe (Germany, France, United Kingdom, Italy, Spain, Finland, Sweden and Switzerland) [6–10]. Until now validation of the Euroscore with a hospital in the Netherlands has not yet been performed.
Most predictive models use perioperative death as an endpoint. Only a few models address other relevant outcomes as morbidity and length of stay on the intensive care unit [7,11–13]. The Euroscore, however, was not developed for predicting postoperative morbidity.
The Amphia Hospital (a teaching hospital at Breda, the Netherlands) has kept an expanding database for more than 10 years with pre, per and postoperative characteristics of all patients who underwent cardiac surgery. Since 1997 the database includes variables of the most commonly used risk-predicting systems, including the Euroscore.
The objective of the present paper is: (1) to define models that predict in-hospital death, major cardiac complications and extended intensive care unit duration for patients who underwent CABG, a heart valve operation or combined; and (2) to validate the Euroscore model in our population.
|
2. Material and methods |
|---|
|
TopAbstract1. Introduction2. Material and methods3. Results4. DiscussionReferences |
|---|
2.2. Data collection
2.2.1. Determinants
We collected information about four clusters of variables relating to demographics, morbidity, cardiac status and indication/intervention. Demographic variables consisted of age, gender and body mass index. The morbidity variables consisted of chronic obstructive pulmonary disease (COPD, defined as a history of COPD with active treatment), preoperative kidney function abnormalities (creatinine level above 150, above 200 µmol/l, or dialysis), diabetes mellitus (types I or II), haemoglobin (expressed in % of lower limit for applicable gender), active smoking (smoking during the week preoperatively), hypercholesterolemia (defined as patients treated actively with cholesterol or lipid lowering drugs), hypertension (blood pressure above 160 systolic or 90 mmHg diastolic or active treatment for hypertension), extra cardiac arteriopathy (defined as documented arteriopathy or a vascular intervention in the history) and neurological dysfunctional disorder (defined as a history of transient ischemic attack or cerebrovascular accident). Cardiac status variables consisted of prior cardiac surgery (none, once, twice or more), prior percutaneous coronary intervention (PCI), angina pectoris class (New York Heart Association (NYHA)), dyspnoea class (NYHA), history of recent myocardial infarction (interval between infarction and cardiothoracic intervention shorter than 1 day, between 1 day and 1 week, between 1 and 4 weeks, older than 4 weeks), left ventricular ejection fraction (good: >50%, moderate: 30–50%, poor: <30%), three vessel disease (50% stenosis of branches of either left coronary descendent, circumflex or right coronary artery), left main stem disease (stenosis >50%), pulmonary hypertension (pre- or peroperative systolic pulmonary artery pressure >50 mmHg) and left ventricular hypertrophy (diastolic left ventricular wall thickness >1.0 cm on ultrasound). And indication/intervention variables consisted of emergency procedure (classified according to the intention to treat the underlying disease either electively, within 1 day of indication, within 1 h of intervention or in cardiopulmonary resuscitation conditions), cardiopulmonary resuscitation before operation, failed PCI (defined as a unsuccessful PCI with the need for same day cardiac surgery), intra-aortic balloon pump (IABP) before operation, acute renal failure (preoperative anuria or oliguria below 20 ml/min for more than 6 h), preoperative use of inotropic or vasopressor medication, the aggregated variable of the Euroscore, ‘critical preoperative state’ [2] and type of operation.
2.2.2. Outcome variables
We collected information about three different outcomes: (1) in-hospital death, defined as death during hospitalisation in the Amphia hospital or in one of the affiliated hospitals; (2) major adverse cardiac events (MACE) defined as in-hospital death or perioperative myocardial infarction (Q-wave or non-Q-wave with creatinin kinase MB fraction (CK-MB)>50 µg/l in the year 1997 or CK-MB>100 µg/l from 1998 onwards) or ventricular tachycardia/fibrillation; and (3) extended length of stay (ELOS) defined as intensive care length of stay of at least 3 days or in-hospital death.
2.3. Analysis
We developed for each of the three outcomes a separate prediction model.
We coded qualitative variables with ‘1’ when the characteristic was present and with ‘0’ otherwise. We categorised the following continuous variables: age (for every 5 years over 60 one point as in the Euroscore), body mass index (<20, 20–24.9, 25–34.9, 35 kg/m2 and higher) and haemoglobin (>90%, 80–90%, <80% of the lower normal limit per gender). We replaced missing values in categorical variables with the most prevalent value. We assessed univariate relations between the determinants and the three different outcome variables by calculating odds-ratios and 95%-confidence intervals for all individual variables. We used multiple logistic regression technique for developing the predicting models. Firstly, all variables were entered per cluster into a stepwise regression analysis. The variable with the smallest P-value was entered into the model. After each entry, variables that were in the model were tested for possible removal based on the likelihood-ratio criterion. Variables that showed a significance level of P-value <0.20 were entered in the model and variables with a P-value >0.20 were removed from the model. Secondly, all remaining variables in the four clusters were entered into a second stepwise logistic regression analysis. Variables with a P-value <0.05 were entered in the model and variables with a P-value >0.10 were removed from the model. We divided the study-cohort at random in two subsets: the derivation set holding two thirds of the entire cohort and the validation set holding the remaining third.
The construction of the Amphiascore involved the following steps: the smallest ß-coefficient of the logistic model was assigned a risk score value of 1; each subsequent ß-coefficient was divided by the smallest ß-coefficient and then rounded to the nearest integer.
2.3.1. Validation and calibration
We calculated the predicted risk for each patient for all three outcome variables by using the developed logistic models. We validated the discriminatory performance of the model internally and externally by receiver-operating characteristics (ROC) curve analysis of the derivation respectively the validation subset.
We compared the observed and predicted event frequencies using the Hosmer-Lemeshow goodness-of-fit statistic [14] for the validation subset. We used the Amphiascore to define five risk categories (low, low-medium, medium, medium, medium-high and high). We chose the thresholds so that the categories contained a similar number of outcome events. We performed this for all three Amphiascores and for the Euroscore.
2.3.2. Comparison with the Euroscore
We compared the Amphiascore and the Euroscore by: (1) summarising the variables in the predicting models with their corresponding odds ratios; (2) by performing ROC curve analysis for both models; and (3) by calculating the amount of agreement between the two scores based on five risk categories by using weighted kappa inter-observer statistics [15]. The following guidelines were used to interpret the strength of agreement: kappa <0.20 was considered to represent poor agreement, 0.21–0.40 fair, 0.41–0.60 moderate, 0.61–0.80 good and 0.81–1.00 almost perfect agreement [16].
2.4. Data management
Data were prospectively gathered from admission in the Amphia-hospital until dismissal or in-hospital death. The entry program contained routines to check for inconsistencies or faulty entries. After completing the pre-, per- and postoperative parts of the information the treating physician was asked to sign for completeness. The database was checked to assure 100% accurateness that: (1) all patients who entered the operating room for a cardio thoracic intervention had an electronic record in the database; and (2) that all patients who died in-hospital were classified as such.
|
3. Results |
|---|
|
TopAbstract1. Introduction2. Material and methods3. Results4. DiscussionReferences |
|---|
Table 1 shows the screened preoperative variables in relation with in-hospital death, MACE and ELOS by univariate analysis. Patients, reaching one of the three outcome points, were generally older (over 70 years), were more often female, had one ore more co-morbidities (most frequent were a high creatinine level or dialysis prior to operation), had significant preoperative cardiac disease (prior cardiac surgery, poor left ventricular ejection fraction or myocardial infarction in the last 7 days), had combined CABG/valve surgery or had one of the following urgency factors: indication, IABP before operation, acute renal failure or inotropic medication.
|
|
|
|
|
Table 5 shows the agreement between the two scoring systems. There is a high agreement in distributing patients to the low (84%) and the high-risk group (59%). The three medium risk groups are less well in agreement: only 26–35% of the patients were classified in the same risk category by both scores. The weighted kappa-statistic, 0.51, shows a moderate level of agreement between the two scores.
|
|
4. Discussion |
|---|
|
TopAbstract1. Introduction2. Material and methods3. Results4. DiscussionReferences |
|---|
4.1. In-hospital death
Our model contains less variables than the Euroscore, but performs as well as. The variable creatinine however was of greater importance in the Amphia hospital than in the Euroscore model. Even patients with a creatinine level above 150 µmol/l had a higher risk of death than was predicted by the Euroscore in patients with a creatinine level above 200 µmol/l. However, patients with a high creatinine level were generally older than patients with a normal creatinine level (58% older than 70 years versus 32%), underwent more prior surgery (16 versus 7%) and more emergency procedures (15 versus 7%). Thus creatinine is an important predicting factor for in-hospital death, but is not independent from other important prognostic factors. Other studies also demonstrated higher risks of death in patients with an abnormal creatinine level or also applied more than one risk level for creatinine [12,18]. Prior surgery was also of greater influence in the Amphiascore than in the Euroscore model. Other studies were in agreement with this observation, especially when considering whether a patient underwent their third (or more) cardiac procedure [1,19]. A myocardial infarction in the last 24 h before the cardiac procedure (0.9% of the patients) was found to be a factor of major impact. Although the Euroscore includes the factor ‘recent myocardial infarction in the last 90 days’, the above-mentioned factor contributed more to our model. Several published models include the factor myocardial infarction in de last 24 h as well in their models [20,21]. Whether the absence of this factor in the Euroscore model limits the model, could be debatable.
4.2. Major adverse cardiac events
Predicting major adverse cardiac events is less accurate than predicting in-hospital death, as we illustrated by the lower level of validation. Besides other thresholds for a preceding myocardial infarction (within 24 h and older than 30 days) the model consisted only of variables that were already in the Euroscore. On the other hand, perioperative variables like increased cross-clamping time, increased operative use of blood products, lower intraoperative diastolic blood pressure and new intraoperative electrocardiographic ST-T changes are factors that have previously been found to be of importance in predicting perioperative myocardial infarction [22–24]. Adding this variables could increase the accuracy of the MACE-predicting model. Because we intended to make a predicting model based on preoperative factors, these factors were not included.
4.3. Extended length of stay on intensive care
The Amphia-ELOS-model was a relatively good prediction model. The developed model included 14 variables, making it the most complicated of the three models. Low haemoglobin preoperatively and failed PCI were, besides the variables from the Euroscore, present in the predicting model.
Overall the Euroscore is a good method to benchmark in-hospital death between the Amphia hospital and other hospitals where the Euroscore was validated. However the weighting of some variables is very different from the Amphiascore and the amount of agreement between the medium risk categories is not optimal. In case of quality assurance within the Amphia hospital itself, the use of the here genuinely developed Amphia models seems quite defendable, certainly in combination with the Euroscore. In the future for quality assurance we will use both scores for comparing standardised mortality ratios between individual surgeons.
4.4. Limitations
The database was intended for both optimising patient care and scientific research. Since we started this database before the Euroscore was published and some variables were not implemented during the whole period, but only in the last 2 years. This could create some bias in the model.
We measured in-hospital death, instead of death within 30 days as used in the Euroscore. This cannot account completely for the observed differences between the developed Amphiascore and the Euroscore. At the other end, patients who died after 30 days in the hospital (12% of the cases) were categorised in the Amphia model as in-hospital deaths, resulting in an overestimation of in-hospital death with respect to the Euroscore.
The two models that are predicting morbidity are not as accurate as the in-hospital death model. A better performing model is difficult, firstly because many peroperative variables play a role in postoperative morbidity, like the extracorporal circulation time, secondly they are a heterogeneous group of outcomes, the length of stay on ICU includes in fact all major complications that can occur after surgery and MACE are also an aggregation of three outcome variables.
Recent myocardial infarction in the Euroscore was defined as a myocardial infarction within 90 days prior to operation. In the Amphia database, time intervals for myocardial infarction were set at 1 day, 1 week and 4 weeks. Since only a minority of the patients (1.3%) who had had a myocardial infarction that occurred more than 4 weeks preoperatively eventually died in-hospital, we do not think this has led to important changes in the model for in-hospital death, and thus will not limit the comparability of the two scoring systems.
The definition of a non-Q-wave myocardial infarction has changed slightly over the years 1997–2001. In 1997 the cut-off point was set at a CK-MB of 50 µg/l, thereafter at 100. We do not believe this to be an important limitation, because the Amphia-MACE model, excluding 1997, resulted in the same variables in the prediction model and only marginally changes in the odds-ratios.
The comparison with the Euroscore model was made using data from only one Dutch hospital. It is then only possible to perform internal validation of the model. We intend to collect data from all Dutch hospitals performing cardiac operations to enable us to use the Euroscore model in the complete Dutch population. In the near future we hope to perform external validation of the model, which will probably optimise the model.
We conclude that the Amphia score performs as well as the Euroscore in discriminating patients with respect to in-hospital death. Our models for predicting major adverse cardiac events and extended length of stay on intensive care may be useful tools in categorising patients in various subgroups of risk for postoperative morbidity.
|
References |
|---|
|
TopAbstract1. Introduction2. Material and methods3. Results4. DiscussionReferences |
|---|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| ANN THORAC SURG | ASIAN CARDIOVASC THORAC ANN | EUR J CARDIOTHORAC SURG |
| J THORAC CARDIOVASC SURG | ICVTS | ALL CTSNet JOURNALS |
