The value of clinical features in the diagnosis of acute pulmonary embolism: systematic review and meta-analysis

Abstract

Background: Clinical assessment of patients with suspected pulmonary embolus (PE) is used to estimate the probability of PE and determine what (if any) diagnostic testing is required.

Aim: We aimed to estimate the diagnostic value of individual clinical features used to determine the pre-test probability of acute PE.

Design: Systematic review and meta-analysis.

Methods: We searched electronic databases (1966 to May 2007) and the bibliographies of retrieved articles for any article that reported the diagnostic performance of clinical features compared to a reference standard diagnostic test in patients with suspected acute pulmonary embolism. Likelihood ratios were calculated for each feature and pooled using a random effects model, as implemented by MetaDiSc statistical software.

Results: We identified 18 studies for inclusion with a total of 5997 patients. The most useful features (pooled likelihood ratio) for ruling in PE were syncope (2.38), shock (4.07), thrombophlebitis (2.20), current DVT (2.05), leg swelling (2.11), sudden dyspnoea (1.83), active cancer (1.74), recent surgery (1.63), haemoptysis (1.62) and leg pain (1.60); while the most useful features for ruling out PE were the absence of sudden dyspnoea (0.430), any dyspnoea (0.521) and tachypnea (0.561). All other clinical features had likelihood ratios near to one. Many of the analyses involved pooling results that had significant heterogeneity, so these estimates should be used with caution.

Conclusions: Individual clinical features only slightly raise or lower the probability of PE. In isolation, they have limited diagnostic value and none can be used to rule in or rule out PE without further testing.

Introduction

Suspected acute pulmonary embolism (PE) is a common cause for acute hospital attendance and admission. Clinical assessment is necessary to estimate a pre-test probability of PE and determine what (if any) diagnostic testing is required. Clinical assessment may be used in an unstructured manner to generate a pre-test estimate of probability or may be used in a formal clinical probability score to categorize patients into (typically) low, intermediate or high-risk groups.¹ In the former approach, a likelihood ratio can provide an estimate of the diagnostic value of each clinical feature and help estimate the clinical probability of PE in a Bayesian manner.² In the latter each valuable clinical feature can be ascribed a score, or weight, that contributes to the overall scoring, or categorization, of clinical probability.

Whichever approach is taken, we need to know how diagnostically useful each element of clinical assessment is. Although there have been many studies examining the diagnostic value of individual clinical features, and meta-analyses have examined clinical probability scores,^1,,3 we are not aware of any studies that have systematically synthesized data to provide an overall estimate of the diagnostic value of each clinical feature used to diagnose PE. It is recognized that diagnostic decision-making should be based upon systematic identification, selection and synthesis of the available literature.⁴

We aimed to systematically review the literature and synthesize existing data to estimate the value of individual clinical features in the diagnosis of acute PE.

Methods

We undertook a systematic review and meta-analysis of diagnostic cohort studies evaluating the value of clinical features for PE in patients presenting with suspected acute PE.

Search strategy

We searched the following electronic databases (1966 to May 2007): Medline, EMBASE, CINAHL, Web of Science, Cochrane Database of Systematic Reviews, Cochrane Controlled Trials Register and Database of Reviews of Effectiveness. The bibliographies of all retrieved articles were scanned for potentially relevant articles that were not identified by the original search.

Selection of studies for inclusion

Two reviewers (J.W. and F.S.) screened the titles and abstracts of all articles identified by the search strategy and independently determined whether the article could potentially be reporting a cohort study that measured the diagnostic value of clinical features or clinical score compared to a reference standard test (pulmonary angiography, lung perfusion scanning (with or without ventilation scanning), CT pulmonary angiography, lower limb ultrasound or D-dimer), although for this study we only ultimately analysed those reporting clinical features. Full copies of all selected articles were retrieved. Two further reviewers (J.W. and S.G.) then independently reviewed the full articles to determine whether they really did meet the criteria outlined above. A Kappa score was calculated for agreement between the two reviewers and disagreements were resolved by discussion.

We specifically excluded studies that measured the risk of developing PE after recording clinical characteristics, rather than the probability of PE being present at the time of assessment; case–control studies, in which patients were selected on the basis of having, or not having, PE; and studies with less than ten patients. We included studies published in English, French and Spanish, but excluded studies published in other languages.

Assessment of study quality

We assessed study quality by determining whether application of the reference standard was independent of the findings of clinical assessment, whether observers who were blind to the reference standard results undertook clinical assessment, and whether observers who were blind to the results of clinical assessment interpreted the reference standard. Empirical evidence suggests that failure to meet these criteria is associated with over-estimation of diagnostic accuracy.⁵

Data extraction

We extracted the following data from each article: the setting for recruitment, groups excluded from the study, population characteristics (mean/median age, gender balance), prevalence of PE, whether clinical data were extracted from clinical notes or collected on a standardized proforma by the clinician, who recorded the clinical data, the reference standard used, and the number of true positives, true negatives, false positives and false negatives for each clinical feature (either as reported or calculated from the reported data).

Data analysis

We used a random effects model, as implemented by MetaDiSc statistical software,⁶ to estimate pooled likelihood ratios for the presence and absence of each clinical feature. A Chi-square test for heterogeneity is reported for each clinical feature.

Results

The flow of articles selected for review is shown in Figure 1. We identified 46 studies examining clinical diagnosis of PE, 28 of these only reported the results of composite clinical scores, leaving 18 that reported individual clinical features.^7–24 The characteristics of the studies included in the review are shown in Table 1. The 18 studies included a total 5997 patients with suspected pulmonary embolism (median sample size 171, range 37–1100), the majority presenting to an emergency department. The prevalence of PE ranged from 9% to 55% (median 38%). The mean age of each study cohort ranged from 44 to 69 years (median 57) and the proportion of male subjects ranged from 30% to 59% (median 41%).

Figure 1.

Flow diagram of studies considered for the review.

Table 1.

Characteristics of the study cohorts

First author	Year	N	% with PE	Mean age	% male	Setting	Reference standard	Independent reference standard	Blind clinical assessment	Blind reference standard	Features assessed (see Tables 2 and 3 for codes)
Hildner FJ7	1967	78	41	NS	NS	IP	Isotope lung scan and PA	Yes	Yes	Unclear	H1, H2, H3, H4, H5, H6, H7, H8, H9, H10, H12, H14, H17, H21, E1, E2, E3, E4, E6
Guidotti TL8	1979	37	35.1	49	40.5	ED	V/Q scan and PA	Yes	Unclear	Unclear	H1, H2, H3, H5, H6, H7, H8, H9, H12, H14, H15, H19, H20, H21, E1, E2, E4
Hull RD9	1988	173	20.8	47	44.5	ED	Venography, impedence plethysmography, V/Q scan and PA	Yes	Yes	Yes	H1, H2, H4, H5, H6, H7, H11, H14, H18, H19, E2, E5
Celi A10	1989	177	54.8	57.7	45.2	ED + IP	LPS	Yes	Yes	Yes	H1, H4, H11, H21, H23, E3
Gray HW11	1990	61	36.1	56	43.6	NS	V/Q scan and PA	Yes	Yes	Yes	H1, H2, H3, H5, H7, H15, H20
Nazeyrollas P12	1996	70	44.3	61.5	38.6	OP	V/Q scan and PA	Yes	Unclear	Unclear	H1, H2, H5, H7, H13, H17
Susec O13	1997	170	15.3	43.7	34.7	ED	V/Q scan and PA	Yes	Yes	Unclear	H1, H2, H3, H4, H5, H6, H7, H8, H9, H10, H11, H16, H20, H22, H23, H24
Grifoni S14	1998	117	53.9	68	37.6	ED	LPS and PA	Yes	Yes	Yes	H1, H2, H3, H4, H5, H6, H11, H10, H13, H14, H15, H16, H18, H19, E6
Leveau P15	2000	249	36.6	69.2	40.6	ED	CUS, phlebography, LPS and PA	Yes	Yes	Unclear	H1, H2, H7, H9, H11, H12, H13, E5
Stollberger C16	2000	168	27.4	65.0	48.1	IP	LPS and PA	No	Unclear	Unclear	H1, H2, H3, H4, H5, H6, H7, H8, H9, H10, H13, E1, E2, E4, E5, E6
Wells PS17	2001	930	9.3	50.5	37.3	ED	D-dimer, CUS, V/Q scan and PA	No	Yes	Yes	H3, H4, H6
Miniati M18	2001	110	39.1	63.8	40.9	IP	LPS and PA	Yes	Yes	Yes	H3, H4, H6, H8, H10, H12, H18, H22
Kline JA19	2002	934	19.4	50.4	30.3	ED	CUS, venogram, V/Q scan, CT and PA	Yes	Unclear	Unclear	H1, H2, H3, H4, H5, H6, H7, H8, H9, H10, H11, H16, H17, H19, H20, H23, H24
Miniati M20	2003	390	41.0	68.5	42.1	IP + other	LPS and PA	Yes	Yes	Yes	H3, H4, H6, H8, H10, H11, H12, H16
Miniati M21	2003	1100	40	NS	45.3	NS	LPS and PA	Yes	Yes	Unclear	H1, H2, H3, H4, H6, H7, H8, H12, H15, H17, H18, H22, H24, E1, E3
Miniati M22	2005	215	43.3	NS	45.6	IP	LPS and PA	Yes	Yes	Yes	H1, H2, H3, H4, H7, H8, H10, H12, H13, E1, E3
Anderson DR23	2005	858	9.6	50.2	35.0	ED	D-dimer, CUS, CTPA and PA	No	Yes	Yes	H3, H4, H5, H6
Hatipoglu O24	2006	160	40.0	57.0	58.8	NS	LPS	Yes	Yes	Unclear	H1, H2, H7, H17

NS = not stated; ED = emergency department; IP = inpatient; OP = outpatient; LPS = lung perfusion scan; V/Q scan = ventilation-perfusion scan; PA = pulmonary angiography; CUS = compression ultrasound (lower limb).

In most studies, the reference standard was applied independently of the results of clinical assessment. The exceptions were studies that augmented a D-dimer reference standard with further testing based on clinical probability. Clinical assessments were blinded in the majority of cases, but reporting of the blinding of the reference standard was generally poor.

Each study reported only a selection of the clinical features. The last column of Table 1 shows which clinical features were reported by each study. The codes for clinical features are outlined in Tables 2 and 3. A few studies reported clinical features in such a way that the relevant data could not be extracted, for example reporting means for continuous data such as tachycardia.

Table 2.

Summary likelihood ratios for features from clinical history

Feature	Code	Number of studies	Feature present			Feature absent
			Likelihood ratio	95% CI	P-value for heterogeneity	Likelihood ratio	95% CI	P-value for heterogeneity
Chest pain	H1	14	1.07	0.855–1.33	<0.001	1.00	0.838–1.19	<0.001
Dyspnoea	H2	13	1.42	1.14–1.78	<0.001	0.521	0.372–0.729	<0.001
Immobilization	H3	13	1.41	1.20–1.60	0.029	0.893	0.856–0.931	0.395
Surgery	H4	13	1.63	1.23–2.12	<0.001	0.897	0.846–0.950	0.019
Active cancer	H6	12	1.74	1.17–2.59	<0.001	0.925	0.871–0.983	<0.001
Haemoptysis	H7	12	1.62	1.23–2.15	0.279	0.983	0.948–1.02	0.016
Past history of venous thromboembolism	H5	10	1.47	1.11–1.96	<0.001	0.940	0.895–0.986	0.650
Oestrogen therapy	H8	9	1.26	0.879–1.81	0.845	0.996	0.987–1.01	0.600
Syncope	H11	8	2.38	1.54–3.69	0.007	0.879	0.790–0.978	<0.001
Trauma	H10	7	1.05	0.575–1.91	0.005	1.01	0.947–1.08	0.004
Cough	H9	7	0.847	1.76–2.76	0.255	0.784	0.721–0.852	0.175
Thrombophlebitis	H12	6	2.20	0.435–3.29	<0.001	0.683	0.399–1.17	<0.001
Current DVT	H13	5	2.05	1.12–3.73	<0.001	0.787	0.650–0.954	0.002
Leg swelling	H17	5	2.11	1.59–2.79	0.089	0.836	0.798–0.876	0.452
Congestive heart failure	H14	4	0.966	0.487–1.92	0.255	1.02	0.917–1.13	0.188
Sudden dyspnoea	H15	4	1.83	1.07–3.13	<0.001	0.430	0.254–0.730	0.001
Smoking	H16	4	0.655	0.527–0.814	0.414	1.17	1.04–1.33	0.020
Cardiovascular disease	H18	4	0.791	0.640–0.977	0.324	1.06	0.928–1.22	0.043
Chronic pulmonary disease	H19	4	0.728	0.533–0.992	0.987	1.06	1.00–1.12	0.768
Pleuritic chest pain	H20	4	0.894	0.745–1.07	0.373	1.05	0.896–1.24	0.200
Leg pain	H21	3	1.60	0.936–2.74	0.783	0.905	0.737–1.11	0.166
Pulmonary disease	H22	3	0.426	0.309–0.588	0.932	1.13	1.09–1.18	0.653
Cardiac disease	H23	3	0.916	0.619–1.22	0.474	1.01	0.917–1.11	0.300
Pregnancy	H24	3	0.398	0.064–2.49	0.011	1.06	0.995–1.14	<0.001

Table 3.

Summary likelihood ratios for features from clinical examination

Feature	Code	Number of studies	Feature present			Feature absent
			Likelihood ratio	95% CI	P-value for heterogeneity	Likelihood ratio	95% CI	P-value for heterogeneity
Tachycardia	E1	5	1.33	0.936–1.88	0.007	0.864	0.715–1.05	0.036
Cyanosis	E2	4	1.343	0.686–2.63	0.885	1.00	0.958–1.05	0.489
Fever	E3	4	0.489	0.207–1.15	<0.001	1.10	0.978–1.24	<0.001
Pulmonary rales	E4	3	1.32	0.904–1.94	0.695	0.926	0.768–1.12	0.298
Tachypnoea	E5	3	1.34	1.15–1.57	0.173	0.561	0.404–0.780	0.312
Shock	E6	3	4.07	1.84–8.96	0.167	0.791	0.649–0.965	0.035

Tables 2 and 3 show the summary likelihood ratios for individual clinical features, along with a 95% confidence interval and a P-value for heterogeneity. The most useful features (likelihood ratio) for ruling in PE are syncope (2.38), shock (4.07), thrombophlebitis (2.20), current DVT (2.05), leg swelling (2.11), sudden dyspnoea (1.83), active cancer (1.74), surgery (1.63), haemoptysis (1.62) and leg pain (1.60); while the most useful features for ruling out PE are the absence of sudden dyspnoea (0.430), any dyspnoea (0.521) and tachypnea (0.561).

Discussion

Clinical assessment should be based upon systematic identification, selection and synthesis of the available literature.⁴ To our knowledge this is the first published meta-analysis of key features of the clinical history and examination used to diagnose PE. We have shown that a history of syncope, thrombophlebitis, current DVT, leg swelling, sudden dyspnoea, active cancer, surgery, haemoptysis or leg pain each slightly increase the likelihood of PE, as does shock on examination. The absence of dyspnoea or tachypnea each slightly reduces the probability of PE. All these features have likelihood ratios relatively near to one, so no clinical feature can be used in isolation to rule in or rule out PE. It is also apparent that a number of clinical features that may commonly be used to diagnose PE, such as pleuritic chest pain does not appear to be supported by the pooled data presented here.

We did not include studies of composite clinical probability scores or studies of overall unstructured clinical estimates of PE in our analysis. Chunilal et al.¹ recently examined these studies and found that clinical probability scores and unstructured probability estimates categorized patients into high, medium and low risk of PE with similar accuracy. High probability estimates by clinical probability score were associated with likelihood ratios ranging from 8.6 to 66 in derivation studies and from 1.4 to 29 in validation studies, while unstructured high probability estimates were associated with likelihood ratios ranging from 1.9 to 12. Low probability estimates by clinical probability score were associated with likelihood ratios ranging from 0.05 to 0.31 in derivation studies and from 0.19 to 0.93 in validation studies, while likelihood ratios for unstructured low probability estimates ranged from 0.13 to 0.53. Derivation studies of clinical scores are likely to overestimate diagnostic accuracy so it seems reasonable to conclude that clinical scores and overall unstructured estimates provide more useful diagnostic assessment than isolated clinical features but are not sufficiently accurate to rule in or rule out PE without further testing.

The clinical probability scores assessed by Chunilal were the Wells score,²⁵ the Geneva score²⁶ and the PISA PED score.²⁷ There is no clear difference in diagnostic performance between the three scores but the Wells score has been more extensively validated. The characteristics used in the scores reflect, to a certain extent, the findings of our analysis. The simplified Wells score uses features of DVT (likelihood ratio 2.05), recent surgery (1.63), haemoptysis (1.62) and cancer (1.74). It also uses tachycardia and past history of DVT/PE, neither of which appeared to be diagnostically useful in our analysis. The Geneva score uses previous DVT/PE, recent surgery (replaced by malignancy in the Modified Geneva score) and tachycardia, alongside age, chest radiograph and arterial blood gas results. The PISA PED score uses sudden dyspnoea (1.83), chest pain (not useful), fever (not useful) and syncope (2.38), alongside age, gender previous cardiovascular, respiratory disease or thrombolphlebitis, electrocardiogram and chest radiograph. The inconsistency among these scores, and between these scores and our findings, suggest that an optimal score has yet to be identified or that characteristics may vary between settings.

Some limitations in our analysis need to be appreciated before using these likelihood ratios in clinical practice. We did not search for unpublished data and studies in languages other than English, French and Spanish were excluded. However, it is unlikely that the poor diagnostic performance of individual features can be attributed to either publication or language bias. A more salient issue is that the primary studies included in our analysis may have used certain clinical features to select patients for inclusion. For example, if patients with pleuritic chest pain or previous venous thromboembolism were preferentially included in the studies over those with non-pleuritic chest pain or no previous venous thromboembolism then this may explain why these features did not appear to be diagnostically useful. Diagnostic studies of patients with suspected PE can only tell us what clinical features increase or decrease the likelihood of PE once PE has been suspected, they cannot tell us whether to suspect PE in the first place.

Another important limitation is that many of our analyses were subject to substantial heterogeneity. This is a common feature of meta-analysis of diagnostic data and there is some controversy as to whether data should be synthesized in the presence of significant heterogeneity. We have chosen to synthesize data because an overall summary estimate provides clinicians with a practical, usable best estimate. However, these summary estimates should be used with caution. The observed heterogeneity is most likely to be due to differences in population characteristics or clinician interpretation of features, so in specific circumstances a clinical feature may be more, or less, useful than our likelihood ratios suggest. Research is required to identify the causes of heterogeneity in estimates of the diagnostic value of clinical features. In particular, are there specific patient groups in which certain clinical features are more, or less, useful?

In conclusion, this meta-analysis has shown that individual clinical features have limited value in the diagnosis of PE. The presence of syncope, current DVT, leg swelling, sudden dyspnoea, active cancer, surgery, haemoptysis, leg pain or shock each slightly increase the probability of PE, while the absence of dyspnoea or tachypnea each slightly reduce the probability of PE.

Acknowledgements

We thank the Information Resources staff in the University of Sheffield School of Health for their help with literature searches.