Although the incidence of pulmonary embolism (PE) increases with age from 1 per 10 000 in childhood to nearly 8 per 1000 in older-aged patients , the prevalence of objectively proven PE in patients with suspected PE is relatively low: between 20% and 30% [2,3]. Computerized tomographic (CT) scanning in suspected PE has a high sensitivity, ranging between 64% and 100%, with a specificity between 89% and 100% ; however, CT scans also have adverse effects, such as an increased lifetime risk of malignancy from radiation exposure and the risk of contrast nephropathy [5,6]. These concerns force physicians to use a diagnostic strategy in suspected PE that results in the lowest possible number of CT scans.
The overall sensitivity of the D-dimer test in detecting PE, using a cut-off value of 500 μg L−1, is estimated at nearly 100%; its reported specificity, however, varies between 30% and 40% [7,8]. Standardized clinical decision rules (CDRs) are therefore additionally used in suspected PE. One frequently used, is the Wells score, which consists of items obtained from history (risk factors for PE), physical examination (such as heart rate and signs of deep venous thrombosis [DVT]) and a subjective item, where the physician can judge whether an alternative diagnosis is more likely than PE . Recently, the simplified version of the Wells score (all items are assigned one point) was found to have a performance similar to that of the original rule in combination with a high sensitive D-dimer test [3,7]. An ‘unlikely’ Wells score combined with a negative D-dimer test safely excludes PE in 20–40% of patients with suspected PE . Of the remaining patients in whom a CT scan should be performed in order to either exclude or confirm PE, only 20–30% have a positive CT scan for PE . CT scanning is thus indicated in the case of a ‘likely’ Wells score or a positive D-dimer test .
It is well known that the diagnostic accuracy of a test can vary with the strength of clinical suspicion . In that case, previously reported accuracy estimates may not be generalizable to all subgroups. We investigated if this is also the case in D-dimer testing in suspected PE. We compared the accuracy of the D-dimer in subgroups of patients with no, one or two items of the Wells score positive.
For this purpose we performed an analysis of data from a prospective multicentre cohort study in patients with clinically suspected PE, reported in detail elsewhere . In brief, the study population consisted of consecutive in- and outpatients in whom acute PE was suspected. The dichotomized Wells score (cut-off value 4), simplified Wells score (cut-off value 1), dichotomized Revised Geneva Score (RGS) (cut-off value 5) and simplified RGS (cut-off value 2) were calculated in each patient and combined with a high-sensitivity quantitative D-dimer test. A CT scan was performed in patients with a ‘likely’ outcome of at least one of the CDRs, or if the D-dimer was elevated (> 500 μg L−1). In the case of an unlikely CDR with all scores and a normal D-dimer result, no CT scan was performed. Patients were followed for 3 months by telephone; they were instructed to return to the hospital if symptoms of PE or DVT occurred and imaging diagnostic tests were carried out if PE or DVT was suspected .
The mean age of the 807 included patients with suspected PE was 52 years; 61% were female. A D-dimer result was available in 723 (90%), of whom 156 patients had PE (22%). Median duration of symptoms was 2 days (interquartile range 1–7); the mean body mass index was 26 (standard deviation 5).
We defined three groups of patients, based on the number of positive items on the Wells score. The first group had none of the Wells items positive. The second group had only one item positive; because our analysis focused on patients with a low pre-test probability, it did not matter which item was positive. Finally, we assessed patients with two positive Wells items, one of which was the subjective item, as it is known that when physicians are aware of a positive D-dimer, they are more likely to consider the subjective item ‘alternative diagnosis less likely than PE’ as positive, regardless of the clinical situation . Consequently, in clinical practice, there sometimes is only one item positive officially, combined with a ‘false positive’ subjective item .
To express the diagnostic accuracy of the D-dimer in each subgroup, we calculated the area under the receiver operating characteristics curve (AUC). To evaluate the significance of difference in the AUC, the z-test statistic was used. We also calculated the sensitivity, specificity, positive and negative predictive values, and positive and negative likelihood ratio of the 500 μg L−1 D-dimer cut-off in each group. All analyses were performed using SPSS software, version 19.0 (SPSS Inc, Chicago, IL, USA). A significance level of 0.05 was used in all analyses.
A total of 177 patients had zero positive items on the Wells score, of whom nine patients had PE (5%). The AUC of the D-dimer test in this subgroup was 0.91 (95% confidence interval (CI) 0.82–0.99) (Fig. 1). Another 300 patients had one item positive: this was PE as the most likely diagnosis in 64%, being immobilized within 4 weeks prior to enrollment in 9%, malignancy in 7%, heart rate above 100 beats per minute in 11%, clinical suspicion of DVT in 3%, and 3% had hemoptysis. In this subgroup of 300 patients, 55 had PE (18%). The AUC of the D-dimer test in this subgroup was significantly lower than in the group with zero items positive: 0.78 (95% CI 0.72–0.84; P = 0.020) (Fig. 1).
A total of 136 patients had two items positive, including the subjective item. Of these patients, 49 had PE (36%). The AUC in this subgroup was also significantly lower than in the subgroup with a zero score: 0.78 (95% CI 0.70–0.86; P = 0.031). The AUC of the D-dimer test was not significantly different between the patients with one and those with two Wells items positive (P = 0.42) (Fig. 1).
The sensitivity of D-dimer at the conventional 500 μg L−1 cut-off in patients with a zero Wells score was 100% (95% CI 69–100). This estimated sensitivity was comparable with that in patients with one (98%, 95% CI 90–100) or two items positive (100%, 95% CI 93–100). However, the specificity was lower with increasing number of items positive: 49% (95% CI 42–75), 30% (95% CI 25–37) and 16% (95% CI 10–25), respectively. In addition, the positive predictive value differed significantly between the no, one and two item groups: 10%, 24% and 33% (P < 0.0001). The negative predictive value did not: 100%, 99% and 100% (P = 0.5).The positive likelihood ratio in patients with no, one and two Wells items positive was 2.0, 1.4 and 1.2, respectively. The negative likelihood ratio, however, was 0.0 for all groups.
Since the accuracy varies, one could consider adapting the D-dimer positivity threshold to make the behaviour similar in subgroups defined by the Wells score. Previous studies have demonstrated that a D-dimer cut-off value, adapted to the clinical probability category of the patient, has a better accuracy in excluding PE than using a single D-dimer cut-off value, regardless of the clinical probability [15–17]. In those studies, the proposed cut-off value was kept at 500 μg L−1 for patients with an intermediate clinical suspicion of PE, but doubled (1000 μg L−1) in patients with a low clinical suspicion, and halved (250 μg L−1) in patients with a high clinical suspicion of PE. In a study by Kabrhel and colleagues, for example, the conventional cut-off of 500 μg L−1 had an overall sensitivity of 94% (95% CI 91–97) at a specificity of 58% (95% CI 56–60), and a NPV of 99.5% (95% CI 99.1–99.7). When probability-dependent cut-offs were used, the overall sensitivity became 88% (95% CI 83–92), specificity was 75% (95% CI 74–76) and NPV was 99.1% (95% CI 98.7–99.4) .
Despite the statistical significance of the differences in accuracy, the subgroups in the current analysis were of moderate size, especially the one of patients with zero positive items, making estimates in subgroups somewhat imprecise. In 84 patients no D-dimer was performed . The reason for these missing D-dimer tests is unclear, but it is possible that these patients had a ‘likely’ result on the Wells rule, in which case a D-dimer test was not necessary to indicate CT scan as the next diagnostic step . As we also included patients with two items positive (including the subjective item counting for 3 points), we have also included patients with a ‘likely’ clinical probability according to the Wells rule in this group of patients .
In conclusion, the diagnostic accuracy of D-dimer testing varies significantly across subgroups defined by the Wells score. In particular, specificity is significantly lower in subgroups with one or more positive items of the Wells score. At the current cut-off value, D-dimer specificity is more than halved in patients with two items positive compared with those with none of the Wells items positive. These findings could be taken into account in improving the diagnostic strategy in patients with suspected PE.