Introduction

Early diagnosis of heart failure is crucial because prompt initiation of treatment can prevent or slow down further progression.^1,2 However, diagnosing heart failure is a challenge, and, when based on presenting clinical features alone, there may remain considerable diagnostic uncertainty.³ Yet, the relative importance of diagnostic information obtained at the bedside and subsequent testing, as well as the optimal diagnostic strategy, is not well determined. This particularly holds for settings where more advanced diagnostic tests, notably echocardiography, are not readily available.

Clinical Perspective on p 2873

The diagnostic value of patient history and physical examination has been questioned,^4,5 albeit studies are few and often have methodological limitations. Importantly, some studies focus on particular diagnostic tests in isolation of other tests and tend to ignore that the diagnostic process always involves multiple sources of information, tests, although not all these tests are explicit (eg, age, comorbidity). Moreover, the selection of patients for diagnostic research can lead to biased estimates of the test's performance when, for example, patients with established heart failure are simply contrasted with subjects without this disease.^6,7 Much attention has been paid to the neurohormone B-type natriuretic peptide (BNP) and the inactive N-terminal counterpart NT-proBNP as diagnostic markers for heart failure, as exemplified by the 2008 European Society of Cardiology (ESC) Heart Failure guideline.⁸ However, although BNP indeed shows excellent diagnostic accuracy, it has mainly been studied without regard to other test results, including signs and symptoms.^9,10 Consequently, the relevance of BNP measurements in patients with suspected heart failure with other clinical information already available is unknown.

We set out to determine the diagnostic value of history, physical examination, and subsequent additional testing including BNP measurements to efficiently and accurately establish a diagnosis of new-onset heart failure in the domain of outpatients presenting with nonacute symptoms.

Methods

Setting and Participants

All patients presenting with symptoms and signs suggestive of heart failure (typically dyspnea, fatigue, signs of fluid retention) were eligible for inclusion in the study. Patients were referred to 1 of the 8 rapid access outpatient clinics by primary care physicians based in the catchment area of 7 participating hospitals throughout The Netherlands. The rapid access outpatient clinics provide a “1-stop shop” diagnostic service to the patient and the referring physician and enabled standardized collection of diagnostic data for the present analyses.

Patients with acute heart failure were excluded because they require immediate therapeutic intervention. Patients with known, established heart failure were also excluded.¹¹

Diagnostic Procedures

The standardized diagnostic work-up included: (1) patient history (including age, gender, dyspnea on exertion, orthopnea, fatigue, edema, medical history, medication, and smoking); 2) physical examination (including palpation of the apex beat, auscultation of the heart, pulmonary and abdominal examination, edema, and estimation of the jugular venous pressure, qualitatively as distension of the external jugular vein; 3) 12-lead ECG; 4) chest x-ray (cardiac-thoracic ratio and pulmonary edema as reported by the local specialist); 5) spirometry (forced expiration volume in 1 second, peak flow, and vital capacity); and 6) standard laboratory assessments (including hemoglobin, electrolytes, C-reactive protein, kidney and liver function parameters). In addition, plasma NT-proBNP was determined using validated commercially available immunoassay kits (Roche Diagnostics GmbH, Mannheim, Germany) and assessed at a central laboratory where plasma samples were stored at −80°C. Finally, echocardiography (M-mode, 2D, and Doppler flow) were performed locally in accordance with the American Society of Echocardiography guidelines.¹² Left ventricular ejection fraction was assessed semiquantitatively. Diastolic function was categorized as normal, impaired relaxation, or restrictive filling by a combination of left ventricular wall thickness, transmitral and pulmonary vein flow patterns, and left atrial volume, according to international guidelines,¹³ with the addition of mitral inflow pattern during Valsalva maneuver or tissue Doppler-derived E' when available, as well as assessing noncardiac explanation of the signs and symptoms in the individual patient, notably pulmonary dysfunction. All images were stored digitally for offline assessments.

Follow-Up

Data on the clinical course during the subsequent 6 months, including the response to therapy, were reported by the referring physician. This follow-up information was collected to provide definitive information on the presence or absence of heart failure for those cases with uncertainty at the time of the initial assessment and was used by the outcome panel during the consensus meetings (see below).

Diagnostic Outcome (Reference Standard)

The primary outcome of the study was the presence of heart failure at the time of initial presentation. Because a uniform reference test is lacking for heart failure, we chose a consensus diagnosis as formal reference standard, in analogy with earlier studies and as recommended by recent diagnostic research guidelines.^7,14,15 An outcome panel judged all the available diagnostic and 6-month follow-up information from each patient to determine the final diagnosis. The result of the NT-proBNP was not available during the consensus meetings, however, because a specific aim of our study was to quantify the added diagnostic value of this test and inclusion of the NT-proBNP result would lead to incorporation bias and overestimate its accuracy.¹⁶ Similarly, we did not assess the added diagnostic value of echocardiography, as it played a crucial role in setting the final diagnosis and has very limited availability in primary care, and because ECG, chest x-ray, and BNP are typically used before considering echocardiography.^17,18

The outcome panel comprised 1 of 4 cardiologists, 1 of 4 pulmonologists, and an outpatient heart failure clinic physician (J.C.K.). The latter member was present at all panel meetings, with a panel always consisting of 1 cardiologist and 1 pulmonologist. The approach to arrive at a final diagnosis was by discussion and then consensus. The panel assessed each patient for the presence of heart failure following the criteria and approach outlined by the ESC.¹¹ For each patient, confirmation of cardiac dysfunction, as well as signs and symptoms suggestive of heart failure, was assessed by the panel. In the event of doubt, the response to treatment during the 6-month follow-up period information was used, as outlined by the ESC guideline. If the panel decided heart failure was present, the potential cause and the presence of systolic and/or diastolic ventricular dysfunction were determined.

Statistical Analysis

The analysis was prespecified as a hierarchical multivariable logistic regression model. The aim was to estimate the contribution of sequential diagnostic information from history, physical examination, and additional testing and to summarize this in a diagnostic algorithm that would allow for optimal estimation of the absolute probability of heart failure being present using a minimal number of patient-burdening or expensive diagnostic tests. We first quantified the univariable association of all variables with the primary outcome. Subsequently, multivariable analysis was done, following the order in which tests are applied in practice and without preselection of variables based on univariable analysis, as this can lead to unstable models.¹⁹ The first model included all variables available from history taking and physical examination. Noncontributing tests were manually (1 by 1) excluded using the likelihood ratio test²⁰ and resulted in model 1. Second-stage models were conditional on model 1 to compute the added value of each of the supplementary tests, ie, ECG, chest x-ray, spirometry, plasma parameters (other than NT-proBNP), and NT-proBNP, resulting in models 2a to 2e. The final stage for our analysis was the assessment of the additional value of a second supplementary test conditional on model 2, resulting in models 3a to 3 d. Tests were considered to provide added diagnostic value if the likelihood ratio test probability value comparing models was <0.05.

The discriminative ability of each model was quantified by the c-statistic and the category-free net reclassification improvement,²¹ which measures reclassification in the right direction, for cases upward and for noncases downward the probability of heart failure scale. The calibration was tested with the Hosmer-Lemeshow statistic.²²

To adjust the coefficients and c-statistic for overoptimistic performance, a shrinkage factor was obtained from a bootstrap analysis (1 000 samples).²³ Finally, a diagnostic rule was derived from the shrunken, rounded, multivariable coefficients to estimate the probability of heart failure presence, ranging from 0% to 100%. Score thresholds for ruling in and ruling out heart failure were introduced based on clinically acceptable probabilities of false-positive (20% and 30%) and false-negative (10% and 20%) diagnoses. The average proportion of missing values was 4%, and we imputed missing values rather than performing a complete case analysis.²⁴ The imputation method used a multiple imputation technique as implemented in SAS (version 9.1.3 [SAS Corp, Cary, NC] and R version 2.11 [http://www.r-project.org]).

External Validation

The final diagnostic rule was externally validated in the datasets from the UKNP²⁵ and Hillingdon²⁶ studies from the United Kingdom. The probability of heart failure was calculated for each patient in the respective datasets applying the diagnostic rule we derived and per dataset the c-statistic was calculated. External validation of the rule was hampered by the fact that not all variables were readily available in the 2 datasets. We only included patients with nonmissing data on the NT-proBNP measurement. For the Hillingdon study validation, we used the subset of patients not assessed in the emergency room. The displaced apex beat was derived from the chest x-ray and defined as a cardiothoracic ratio being >60. The BNP was converted to NT-proBNP by means of the formula given by Alibay et al: ₁₀log(NT-proBNP)=1.1*₁₀log(BNP)+0.57; units: pg/mL.²⁷ The UKNP study lacked the “heart murmur suggestive of mitral regurgitation” variable, for which we used the echocardiographic mitral valve regurgitation to impute, with a penalty of scoring only 5 points instead of 10.

Results

The characteristics of the 721 patients enrolled are shown in Table 1 and Figure 1; 6 patients refused to participate (Figure 2). Patients were on average 70.7 years of age (64.6% female). Of the patients, 51.7% had hypertension, 26.1% had chronic obstructive pulmonary disease, and 15.3% had diabetes mellitus. All patients had 1 or more complaints compatible with heart failure, notably dyspnea on exertion, ankle swelling, fatigue, or orthopnea. Of the patients, 7.8% had elevated jugular venous pressure, 9.7% a displaced apex beat, and 27.3% bilateral ankle swelling. In 10 patients, a third heart sound was heard.

• Download figure
• Download PowerPoint

• Download figure
• Download PowerPoint

A final diagnosis of heart failure was made in 207 patients (prevalence 28.7%; 95% CI 25.4–32.2), of whom 72 (34.8%) had a normal systolic function. Follow-up information on the clinical course from the general practitioner at 6 months was unavailable for 12 patients; these patients were contacted individually. Table 1 also shows the univariable associations of all potential diagnostic predictors.

Age, history of coronary artery disease, use of a loop diuretic, and 6 items from the physical examination (displaced apex beat, irregular pulse, rales, pulse rate, heart murmur suggestive of mitral regurgitation, elevated jugular venous pressure) were independently associated with the presence of heart failure. Without any further testing, this yielded a c-statistic of 0.83 (model 1, Table 2).

The only standard laboratory test (ie, excluding NT-proBNP) with sufficient added diagnostic value was “GGT >2 times upper limit of normal” (model 2a). The addition of chest x-ray (cardiothoracic ratio and pulmonary vascular redistribution, model 2b) to model 1 also contributed to the diagnosis, as did the single ECG item “normal” ECG (model 2c) and spirometry (model 2d). The largest added value was provided by the log-transformed NT-proBNP (model 2e), reaching a c-statistic of 0.86, which translates to a net reclassification improvement of 69%, ie, the sum of 21% (net) of patients with heart failure having a higher probability of heart failure plus 48% (net) of patients without heart failure having a lower probability of heart failure. When applying a 20% threshold below which heart failure is considered excluded (for the time being) and 70% above which heart failure is considered present, then the net reclassification improvement for these three categories is 17%, ie, 13% (net) of patients with heart failure move up at least one category, whereas 4% (net) of patients without heart failure move down at least one category. In addition to the model including signs and symptoms and NT-proBNP (model 2e), the chest x-ray conferred the largest diagnostic improvement (model 3b). Of note, the addition of ECG to model 3b did not increase the diagnostic accuracy (data not shown).

All multivariable statistical models were found to have adequate calibration, with probability values from the Hosmer-Lemeshow test all >0.3 and chi-squares all <9.

We recalibrated model 2e with the computed shrinkage factor of 0.92 to account for overfitting and finally rounded the coefficients to formulate the diagnostic rule (Table 3, c-statistic 0.85). The summed score can be transformed into the absolute probability of the presence of heart failure by reading from Figure 2.

The use of the diagnostic rule in 105 patients, of whom 29 had heart failure in the Hillingdon study, resulted in a c-statistic of 0.950. The use of the diagnostic rule in 302 patients from the UKNP study, by means of external validation, resulted in a c-statistic of 0.884. Even when all these patients were set at “no heart murmur suggestive of mitral regurgitation,” the c-statistic was 0.882.

Discussion

The findings among a large unselected group of patients suspected of having new-onset nonacute heart failure in primary care shows that a history taking and physical examination will markedly reduce diagnostic uncertainty. Of the potential additional tests, NT-proBNP yielded the highest additional diagnostic value. The validation datasets confirmed these findings (Figure 3).

• Download figure
• Download PowerPoint

To appreciate these results, some aspects of the study need to be addressed. Previous reports on the diagnostic power of the physical examination of suspected heart failure are scarce. Warnings have even been given for the serious limitations of physical examination,³ and very low univariable positive predictive values of solitary physical examination items have been published in relatively small studies.²⁸ In the present study, however, pulse rate, pulse regularity, pulmonary rales, displaced apex beat, heart murmur suggestive of mitral regurgitation, and distended jugular veins were all univariably associated with the presence of heart failure and retained that association in the multivariable analyses. We decided to exclude the third heart sound from our multivariable models mainly due to its low prevalence (10 patients). By itself, however, it holds a positive predictive value of 0.90, making it a specific marker of heart failure. When using only the history taking and physical examination findings in the validation datasets, the c-statistic reached 0.87 in both the Hillingdon and UKNP study (data not shown), albeit incorporation bias cannot be ruled out in all 3 studies. One could reasonably argue that it is important for diagnosticians to maintain their physical examination skills.

Owing to the large study size, we were able to assess the value of a range of tests potentially associated with the presence of heart failure in multivariable analyses, thereby following the natural hierarchy of daily clinical practice. For example, the chest x-ray contributed to the diagnostic accuracy when added to the clinical assessment, also in case NT-proBNP is known (model 3b), which concords with a study by Cowie et al²⁶ that reported a (statistically nonsignificant) multivariable odds ratio of 2.65 for the cardiothoracic ratio. Badgett et al reviewed 29 studies and concluded that chest x-ray alone cannot adequately exclude or confirm left ventricular dysfunction in clinical settings.²⁹ According to several guidelines, a normal ECG rules out heart failure, based on the high negative predictive value (univariable) for echocardiographic systolic dysfunction of 98%, studied in a population with 18% systolic dysfunction and 51% normal ECG.³⁰ Others have also shown that a normal ECG may be useful to exclude heart failure.^5,31 In the present study, the negative predictive value was 94% in a population with 24% systolic dysfunction and only 27% normal ECGs. In our study, an ECG added significant diagnostic accuracy to readily available diagnostic tests (model 2c) even when NT-proBNP was included in the model (3c). In contrast, one of the few diagnostic studies in this field that performed a multivariable analysis concluded that the ECG did not contribute to a diagnostic assessment including NT-proBNP.²⁵

Knowledge of the plasma level of NT-proBNP resulted in the largest gain in diagnostic accuracy. Three other studies of similar patients comparing BNP with other readily available diagnostic tests concluded that BNP is the most powerful diagnostic test.^25,26,32 The vast majority of other studies assessing the diagnostic value of NT-proBNP were, however, performed in hospital emergency rooms and not in an outpatient or primary care setting. It should be emphasized that univariable diagnostic characteristics, such as sensitivity and specificity, cannot be routinely generalized from 1 setting to another.³³

We selected the diagnostic rule including NT-proBNP as an additional test; it adds considerable diagnostic power, blood can be drawn at the patient's home, and the measurement can be performed in most laboratories. In addition, point-of-care tests for NT-proBNP are available, further facilitating the application of the diagnostic rule. Because there were only gradual differences with other diagnostic models (see Table 2, for example, one including ECG or chest x-ray instead of NT-proBNP), alternative strategies exist should NT-proBNP measurement not be available. We believe our study population is representative of, and our study results generalizable to, the large population of patients with suspected heart failure, ie, a population characterized by a mean age of approximately 70 years, predominantly white and female, often with considerable comorbidity and receiving polypharmaceutical therapy with a mixture of signs and symptoms compatible with heart failure. The prevalence of heart failure was 28.7%, which is similar to comparable previous studies.^25,26,31,32 The use of a loop diuretic was a predictor of the presence of heart failure in our study, which seems somewhat surprising as its prescription is often related to heart failure, but is understandable from a clinical point of view. Patients presenting with symptoms of volume overload will often be treated with loop diuretics to alleviate symptoms, usually before tests to confirm heart failure are performed.

In the present study, we did not assess observer variability within a panel regarding the classification of heart failure as present or absent. In another study from our group, a comparable method was used, and a sample of 41 cases was reassessed resulting in disagreement in only 1 diagnosis.¹⁴ The outcome panel used all available information excluding the NT-proBNP values. Incorporation bias cannot be excluded for other variables.

Finally, we analyzed a substantial number of diagnostic tests, and although we adjusted for overfitting and optimism as much as possible, the accuracy of our rule should be further assessed by means of external validation studies. The external validation in 2 independent external datasets demonstrates the robustness of the diagnostic rule.

For daily clinical practice, the estimation of the probability of the presence of heart failure using the point score estimated in this analysis may help to make decision making more explicit. Consider thresholds whereby a disease probability of <10% (13 points) is considered to rule out heart failure and a probability of >70% (54 points) as ruling in heart failure (Table 4). Application would have resulted in 96 (13.3%) patients assumed to have heart failure and 233 (32.3%) patients assumed not to have heart failure, leaving 392 (54.4%) patients in the intermediate no-decision group. The false-positive and false-negative rates are 12.5% (12 of 96) and 4.7% (11 of 233), respectively.

In conclusion, the findings obtained in this study conducted in a large unselected group of patients suspected of having new-onset heart failure in primary care support the view that history taking and, in particular, physical examination are major sources of diagnostic information. The estimation of their diagnostic contribution will help to improve clinical decision making. The largest additional diagnostic contribution is provided by measurement of NT-proBNP. For daily practice, these findings as well as the diagnostic rule derived from them, may be useful to quantify the probability of heart failure in patients with new symptoms suggestive of heart failure.

Sources of Funding

This study was funded by the Dutch Ministry of Health, ZON-MW grant 945-02-014. Roche Diagnostics supplied the kits for the NT-proBNP assessments. The sponsors of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

Disclosures

None.

Footnotes