Between January 2003 and June 2004, 2,273 primary care patients from 58 practices in the city of Göttingen (North Germany) and the surrounding communities were invited by their general practitioners to participate in the study. Inclusion criteria were the presence of at least one cardiovascular risk factor documented by the family physician, including arterial hypertension, diabetes, family history of early heart disease, and coronary artery disease. Patients were classified as hypertensive if hypertension was documented by their treating physician or if they were on antihypertensive therapy. Patients were classified as diabetic if this diagnosis was made by their treating physician or if they were on antihyperglycemic therapy.

Exclusion criteria were the diagnosis of heart failure documented by the family physician or the presence of a terminal or disabling chronic disease. Patients received a leaflet with general information about the purpose of the study and how to participate. A total of 542 patients agreed to participate and were examined at the Georg-August-University, department of cardiology. Patients were interviewed, clinically examined, and assessed by echocardiography. Fifteen ml of blood were taken from a forearm vein for the measurement of NT-proBNP level.

Before the study appointment, patients completed an overnight fast except for taking their regularly prescribed medications. A 21-gauge butterfly needle was inserted intravenously in the forearm, and after a 30-minute supine rest, blood samples were drawn into lithium-heparinate tubes; these were centrifugated and the supernatant was divided into aliquots and stored at minus 70°C. We used the Elecsys^® assay (Roche Diagnostics, Mannheim, Germany) to measure NT-proBNP in defrosted samples. The lowest detectable measurement for this assay was 5 pg/mL. The interassay coefficient of variation was 2.7% for 175 pg/mL, 2.7% for 355 pg/mL, 1.9% for 1068 pg/mL and 1.8% for 4962 pg/ml. The laboratory technician who measured NT-proBNP levels was at a different site and blinded to the characteristics of the patients and the results of echocardiograms. NT-proBNP reagents were kindly provided by Roche Diagnostics (Professor G. Hess, Mannheim, Germany).

Echocardiograms were performed with a Phillips Ultrasound System (Phillips Sonos Agilent) using a 3.5-MHz transducer with the patient in the left lateral position. A complete resting 2D echocardiogram and Doppler ultrasound examination was performed. We obtained standard 2D parasternal long and short-axis to determine left ventricular dimensions. We calculated left ventricular ejection fraction (EF) by the quantitative 2-D (biplane Simpson) method. An EF<50% was defined as systolic ventricular dysfunction. Echocardiograms were additionally rated concerning abnormalities defined as diastolic dysfunction 10. Three of the authors (CL, SK, RW) interpreted all echocardiograms and were blinded to the results of the NT-proBNP assay as well as to details of the medical history.

Assessing group differences, we used t-tests for continuous variables and Chi-Square-tests for comparisons of frequencies. A multiple logistic regression analysis was done to identify clinical variables that have a statistically significant diagnostic value in predicting left ventricular systolic dysfunction. We used a backward conditional model, including sociodemographic and clinical variables. Selected variables were used to estimate an individual patient risk score, thus, the sum of the β coefficients for each of the specific risk factors multiplied by their actual values 11.

We assessed the diagnostic performance of the NT-proBNP assay and the risk score by using receiver operating characteristic curves (ROC). The overall discriminative power of NT-proBNP and the clinical risk score is shown by the area under the curve (AUC). Comparisons between AUCs were assessed according to the method of Hanley & McNeil 12. In addition to the AUC, we compared the test accuracy of NT-proBNP and the clinical risk score using the McNemar test (i.e. comparison of discordant pairs of false classifications). We chose cut-off points that gave comparable high sensitivity levels and moderate (at least 40%) specificity levels in order to optimize the negative predictive power of the test (SnNout; very high sensitivity: negative result rules out the diagnosis/disease) 1314.

All analyses were two-tailed and the alpha was defined at 0.05. Statistical analyses were carried out using SPSS and Microsoft EXCEL. The study was approved by the local ethics committee, and all patients gave written informed consent before examination.

Table 1 shows the sociodemographic and clinical characteristics of the 542 patients. In 23 patients (4%), EF was below 50%. These patients were significantly older, and complained significantly more often about dyspnea and ankle swelling than those with an EF ≥ 50%. Patients with a reduced EF also reported more often a history of myocardial infarction, coronary artery disease, and hyperlipidemia, and more often took diuretics and lipid-lowering agents. Levels of NT-proBNP were significantly higher in patients with systolic dysfunction when compared to patients with preserved left ventricular function. Three hundred and ninety nine patients of those without systolic dysfunction showed signs of diastolic dysfunction; in most cases (83%) of the lowest grade, i.e. impaired relaxation.

A logistic stepwise regression with left ventricular dysfunction as the dependent variable and 17 independent variables as covariates was conducted. The backward conditional model selected the following three covariates: dyspnea at exertion combined with ankle swelling, history of coronary artery disease, and treatment with diuretics. Other variables were entered into the model but were excluded by the stepwise regression: age (dichotomized by the median), sex, BMI>30, diabetes mellitus, hypertension, hyperlipidemia, myocardial infarction, family history of early heart disease, treatment with ACE inhibitors, β blockers, calcium channel blockers, AT1 blockers, or lipid lowering agents (Table 2).

Figure 1a shows the AUC for NT-proBNP. This illustrates its diagnostic value for left ventricular systolic dysfunction, which can be rated as excellent. The cut-off point of NT-proBNP was determined as having a moderate specificity and a high sensitivity compared to the clinical risk score. Concentrations below the cut-off point allowed for the exclusion of left ventricular systolic dysfunction. The negative likelihood ratio of a NT-proBNP value below 98.5 pg/ml is 0.19. It indicates that the probability of a negative test result is five times lower in patients with impaired left ventricular systolic dysfunction than in patients with preserved left ventricular function. Table 3 shows the cut-off point dependent measures of NT-proBNP as a diagnostic test.

Figure 1b shows the AUC for the clinical risk score including the following variables: suffering from ankle swelling and dyspnea at exertion, history of coronary artery disease, and treatment with diuretics. The backward conditional model with left ventricular dysfunction as the dependent variable selected these risk factors (Table 2). To obtain the individual patient risk score, the β coefficients for each of the specific risk factors were multiplied by their actual values and then summed up. For example, individual patient risk score = (β₁ × dyspnea at exertion × ankle swelling) + (β₂ × coronary artery disease) + (β₃ × diuretics). AUC of the clinical risk score indicates its diagnostic value for left ventricular systolic dysfunction and can be rated as excellent.

The cut-off point was determined as having a moderate specificity and a high sensitivity compared to the NT-proBNP test. The negative likelihood ratio of a clinical risk score below 1.11 is 0.14 and indicates that the probability of a negative test result is seven times lower in patients with impaired left ventricular systolic dysfunction than in patients with preserved left ventricular function. Table 3 shows the cut-off point dependent measures for the clinical risk score as a diagnostic test.

The minimal difference of 0.019 between AUCs of NT-proBNP and the clinical risk score was not significant (Figure 1). Table 4 shows the number of correct and false classifications for both, the risk score and NT-proBNP as single diagnostic tests. Comparing the two tests, there are four different cases possible: Cases shown in the first line and first row (n = 203) were correctly classified by both tests, i.e. they are true positive and true negative diagnoses. Cases shown in the second line and second row (n= 136) were falsely classified by both tests, i.e. they are false positive and false negative diagnoses. Cases shown in the first line and second row (n = 148) were falsely classified by NT-proBNP but correctly classified by the risk score. Cases shown in the second line and first row (n = 55) were falsely classified by the risk score but correctly classified by NT-proBNP. The McNemar test showed that this difference concerning the discordant pairs of false classifications was significant. The advantage of using the clinical risk score for classification was based on the better specificity, resulting in a lower rate of false positive cases.