Introduction
Venous thromboembolism (VTE) is a common, potentially life-threatening condition [1-3]. The burden of disease remains high and presents a leading contributor to the disability-adjusted-life-years loss worldwide [4]. Therefore, prompt and reliable diagnosis of VTE is crucial, as a false positive diagnostic test can lead to unnecessary imaging studies and anticoagulation treatment and is associated with risks such as bleeding, radiation exposure, increased health care costs, and patient inconvenience [3]. On the other hand, a false negative test may lead to increased mortality and morbidity—such as secondary pulmonary embolism (PE), untreated thromboembolic pulmonary hypertension, or post-thrombotic syndrome [3].
American and European guidelines propose a diagnostic approach involving D-dimer measurement as an exclusion test for patients with low or intermediate pretest probability (PTP), as evaluated by a clinical decision rule, and radiological evaluation in the case of pathological D-dimer or high PTP [3, 5, 6].
D-dimer molecules are produced in the context of fibrinolysis, when cross-linked fibrin is degraded [7, 8]. However, even minor activation of the coagulation cascade involves increased fibrinolysis and consequently leads to raised D-dimer levels [9] and, thus, can cause false positive D-dimer results when patients are assessed for VTE [7, 10]. Such confounding factors lead to a significant number of potentially unnecessary radiological diagnostics [3].
Therefore, many attempts were undertaken to optimize D-dimer cut-off values taking into account patient-specific factors such as age, PTP, renal function, or pregnancy [7, 11-15]. However, utility and limitations of suggested D-dimer adjustment strategies were not comprehensively compared.
The aim of this systematic review and meta-analysis is to systematically assess test performance of patient-adapted D-dimer cut-offs for diagnosis of VTE.
Methods
This systematic review and meta-analysis were conducted in alignment with the Cochrane Collaboration guidelines for systematic reviews [16]. The protocol was registered on PROSPERO (No. CRD42020144666, submitted November 1st, 2019). The PRISMA (preferred reporting items for systematic reviews) [17] and MOOSE (meta-analysis of observational studies in epidemiology) [18] guidelines were adhered to for data extraction, quality assessment, and reporting.
Data sources and searches
We performed a systematic search on PubMed, Embase, ClinicalTrials.gov, and the Cochrane Library databases for articles published until 31 December 2021. Furthermore, we systematically searched the bibliographies of eligible publications, references of identified reviews, case reports, and editorials for further relevant studies. No restrictions in respect to region, language, or date were applied. The search strategy was tailored to the specificities of each database but generally followed the same Boolean logic including generic terms for “D-dimer” combined with generic terms for “adapted cut-offs” and “thromboembolism” (see Fig. S1).
Eligibility criteria
All articles investigating patient- or disease-adapted D-dimer levels for VTE were eligible. The exclusion criteria were (i) studies that evaluate D-dimer measurement for purposes other than VTE exclusion; (ii) studies that investigate D-dimer solely as an inclusion test; (iii) studies with high PTP for VTE (≥50%) [3, 5, 6]; (iv) studies with insufficient description of test characteristics to extract; (v) studies investigating D-dimer to guide termination of anticoagulation or duration of treatment or to assess the risk of VTE recurrence or follow-up after VTE; (vi) studies that compare different D-dimer assays or clinical decision rules; (vii) studies including children (<16 year of age); and (vii) animal/experimental trials, or articles that do not provide new original data (narrative reviews, journal clubs, editorials, case reports, etc.). In addition, investigations of subpopulations of previously published trials were excluded.
Definitions
Venous thromboembolism (VTE)
VTE was defined as either deep venous thrombosis (DVT) or PE [3].
Standard D-dimer assessment
A conventional or unadjusted D-dimer cut-off value was defined as a D-dimer cut-off of 500µg/L as per current guidelines [3].
Patient-adjusted D-dimer cut-offs
Adjustment was defined as any type of adaptation of D-dimer cut-off to a patient-specific condition (e.g., age, PTP, renal function, and pregnancy).
Safety and efficiency
Safety and efficiency were defined in-line with previous publications [19-22]. Safety was defined by projected failure rate [21], whereas efficiency refers to the potential number of imaging studies saved.
Study outcomes
The primary objective of this study was to assess and compare test performance of different adjustment strategies of D-dimer threshold to patient-specific factors. Furthermore, we aimed at comparing mortality, ICU-admission rate, hospitalization rate, necessity and length of hospitalization and necessity of imaging studies when patient-adjusted D-dimer cut-offs were used. However, after completion of data extraction, it became obvious that the required information was only presented in the minority of trials. Hence, we performed a sensitivity analysis exploring false negatives, projected failure rate, and false positives at different pretest probabilities (low and intermediate) in order to explore the safety and efficiency aspects of patient-adjusted D-dimer cut-offs.
Study selection, data collection, and data extraction
Three consecutive steps were undertaken by two independent investigators (TK, JLG) to identify eligible articles: (i) While applying predefined eligibility criteria, the titles and abstracts were evaluated for study inclusion. In the case of an unequivocal criteria violation, the study was excluded. If insufficient data were available which hampered decision-making at this time point, the article was considered for further assessment. Even though identified reviews, editorials/comments, and case reports were excluded from our investigation, their bibliographies were screened for further articles. (ii) While applying the predefined inclusion and exclusion criteria, the full-text of identified articles was assessed and their bibliographies screened for potential further investigations. If a full-text was not accessible, authors were contacted twice and then excluded. (iii) Lastly, after the consideration of identified articles for data extraction, data were retrieved by two investigators (TK, JLG) independently using a predefined spreadsheet, as recommended by the Cochrane handbook [16, 23]. Generally, all available data on D-dimer patient-adjustment strategies were extracted from each identified publication. Thus, if an investigation evaluated more than one D-dimer patient-adjustment strategy (i.e., age and YEARS), data for both strategies were extracted. At each step, the decision taken by two independent investigators (TK, JLG) was compared and potential discrepancies resolved by discussion within the study team (CAP, ASM, MM).
Risk of bias assessment
Methodological quality of included studies was assessed by using the Quality Assessment of Diagnostic Accuracy Studies 2 score (QUADAS-2) [24, 25] as proposed by the Cochrane collaboration [23]. Two independent assessors (TK, JLG) individually assessed the risk of bias and methodological quality of each trial rating the risk of bias as “low risk of bias,” “unclear,” or “high risk of bias” based on the QUADAS-2 scoring system. The following risks of bias were evaluated based on the QUADAS-2 score: (i) patient selection, (ii) use of an index test, (iii) use of a reference standard, and (iv) performance and detection bias.
Statistical analysis
The statistical analysis was performed with Stata 16.1 (StataCorp LLC, College Station, TX, USA) and its midas (mainly) and metandi (see later) command [26]. For each subgroup, the following measures/tests were performed: Out of the extracted 2×2 table of each study, pooled values of sensitivity, specificity (i.e., positive and negative predictive value [PPV, NPV], and positive and negative likelihood ratio [LR+, LR−]) were obtained using a bivariate model. A hierarchical logistic regression model (hierarchical summary receiver operating characteristic, HSROC) was constructed using Stata's metandi command to assess the thromboembolism diagnostic accuracy [26]. We assessed publication bias with a funnel plot; the asymmetry was tested using Deek's test. I2 statistics and chi-square test were calculated (two values greater than 50% were considered to indicate substantial heterogeneity) to evaluate heterogeneity between studies [27].
Results
A total of 7549 articles were identified after duplicate removal. Following article screening, we retrieved 142 (1.9%) studies for full-text assessment, and, finally, 68 (0.9%) studies involving 141,880 patients were included in this systematic review and meta-analysis. The PRISMA flowchart is depicted in Fig.1.
We identified studies adjusting D-dimer threshold for age, PTP, YEARS algorithm, COVID-19, renal function, and pregnancy (see also Table S2 for detailed information of each adjustment method). Sixty-six (97.1%) studies were observational trials, one (1.4%) trial was randomized-controlled, and one (1.4%) was an individual patient data meta-analysis. Included and excluded studies are presented in Tables S1 and S3, respectively. Overall risk of bias in included studies was moderate (see Fig.2 and QUADAS-2 assessment in Table S4).
Performance of standard and adjusted cut-offs
HSROC curves and test characteristics for standard cut-off, age- and PTP-, and COVID-19-adjusted D-dimer cut-offs are shown in Fig.3. In addition, the figure shows test characteristics for the most commonly used scores: “Conventional” age-adjusted cut-off (“age×10µg/L above 50 years” [“age×10”]) and YEARS score (combination of PTP assessment and clinical decision rule). Test characteristics of renal function- and pregnancy-adapted cut-offs are listed in Table1, as the numbers of studies for these categories were not sufficient for meta-analysis. HSROC curves for prospective versus retrospective studies, PE versus DVT, and PTP are shown in Fig. S3.
| First author | Year | n | Type of VTE | Prevalence (%) | Sensitivity | Specificity | NPV | Ref |
|---|---|---|---|---|---|---|---|---|
| Renal function-adjusted cut-offs | ||||||||
| Lindner | 2014 | 238 | PE | 16.0 | 0.95 | 0.14 | 0.93 | 28 |
| Pfortmueller | 2016 | 9716 | VTE | 3.4 | 0.86 | 0.56 | 0.99 | 29 |
| Schefold | 2020 | 14,477 | VTE | 3.5 | 0.90 | 0.57 | 0.99 | 7 |
| Pregnancy-specific cut-offs | ||||||||
| Goodacre | 2019 | 168 | PE | 25.6 | 0.70 | 0.33 | 0.76 | 30 |
| Van Der Pola | 2019 | 498 | PE | 4.0 | 1.00 | 0.92 | 1.00 | 14 |
| Zhang | 2021 | 289 | PE | 11.4 | 1.00 | 0.35 | 1.00 | 31 |
- a YEARS algorithm with minor modification VTE=venous thromboembolism, PE=pulmonary embolism, NPV=negative predictive value.
Cumulative prevalence of VTE was comparable between the unadjusted (7.7%) and across adjusted groups (7.9%). Ten of 68 studies did not provide sufficient data for the standard cut-off. Bivariate HSROC evaluation of the standard cut-off revealed a sensitivity of 0.99 (95% confidence interval 0.98–0.99) and specificity of 0.23 (95% CI 0.16–0.31) with an AUC of 0.92 (95% CI 0.89–0.94). Sensitivity was comparable to the standard cut-off for age-adjusted cut-offs (0.97 [95% CI 0.96–0.98]) as well as for “age × 10” (0.97, [95% CI 0.96–0.98]), and YEARS algorithm (0.98, [95% CI 0.91–1.00]) but somewhat lower for PTP-adjusted (0.95, [95% CI 0.89–0.98]), and COVID-19-adapted thresholds (0.93, [95% CI 0.82–0.98]). Specificity was significantly higher for all adjustment strategies (age-adjusted: 0.43, [95% CI 0.36–0.50]; PTP: 0.63, [95% CI 0.51–0.73]; YEARS algorithm: 0.65, [95% CI 0.39–0.84; and COVID-19: 0.51, [95% CI 0.40–0.63]).
Deek's funnel plot asymmetry test, as presented in Fig. S2, suggested significant publication bias for standard (p<0.01) and age-adjusted cut-offs (p=0.01 and p=0.03 for “age × 10”), but not for other adjustment strategies. Heterogeneity and inconsistency amongst studies were high across all subgroups (age-adjusted: I2=100%, 99%–100% and Cochran's Q p<0.001, PTP: I2=99%, 98%–99% and Cochran's Q p<0.001; YEARS: I2=99%, 99%–100% and Cochran's Q p<0.001; COVID-19: I2=87%, 74%–100% and Cochran's Q p<0.001).
Comparison of patient-adjusted cut-offs
Among different adjustment strategies, the YEARS algorithm provided the best negative likelihood ratio (0.03, [95% CI 0.01–0.15]), followed by age-adjusted (both 0.07, [95% CI 0.05–0.09]), PTP-adjusted (0.08, [95% CI 0.04–0.17]), and COVID-19-adjusted thresholds (0.13, [95% CI 0.05–0.32]). Regarding positive likelihood ratio, the YEARS algorithm (2.8, [95% CI 1.4–5.4]) and PTP-adjusted cut-offs (2.6, [95% CI 1.9–3.4]) seemed superior to adjustments to COVID-19 (1.9, [95% CI 1.6–2.4] and age (1.7, [95% CI 1.5–1.9]). Diagnostic odds ratio, as a measure of effectiveness, was highest for the YEARS algorithm (100, [95% CI 14–745]), followed by PTP-adapted (32, [95% CI 15–71]) and age-adjusted cut-offs (25, [95% CI 17–36]), and lowest for COVID-19-adapted thresholds (15, [95% CI 6–37]).
Sensitivity analysis: Robustness, safety, and efficiency according to PTP
Based on HSROC analyses, we calculated test performances and projected failure rates of adjustment strategies and the standard cut-off in an exemplary population of 1000 patients with VTE prevalence of 5% (low PTP) and 20% (intermediate PTP) [3]. Results are shown in Fig.4. Overall projected failure rate was 0.4% (95% CI 0.3%–1.0%) for the standard and 0.5% (95% CI 0.5%–1.0%) for all patient-adjusted D-dimer cut-offs, whereas the potential number of imaging studies saved was 223 (95% CI 167–287) out of 1000 patients. Variability was substantial among the different subgroups.
Discussion
This systematic review and meta-analysis of 68 studies involving 141,880 patients suggest that patient-adjusted D-dimer cut-offs are comparable in test performance to the standard D-dimer cut-off of 500µg/L. Sensitivity of patient-adjusted cut-offs was only slightly lower, whereas specificity increased significantly, and AUC stayed comparable. Optimized test accuracy implies a substantial potential for reduction of imaging procedures when patient-adjusted D-dimer cut-offs are used, while maintaining safety in the sense of a low failure rate.
D-dimer measurement for exclusion of VTE remains a trade-off between optimal safety and test efficiency. The well-established standard cut-off is optimized regarding safety, while coming at the cost of a considerable false positive rate, particularly in the presence of confounders. Our study implies that patient-adjusted D-dimer threshold strategies for the diagnosis of VTE are potentially safe, and it is feasible to rule out VTE. NPV remained similar to CT pulmonary angiogram [32] or superior to compression ultrasound [32] with patient-adjusted D-dimer cut-offs. Our findings go in-line with a recently published randomized controlled clinical non-inferiority trial on safety of using YEARS-algorithm combined with an age-adjusted D-dimer cut-off versus an age-adjusted D-dimer cut-off alone in 1414 patients with suspected PE [20]. At 3 months, the study revealed a missed VTE rate of 0.15% and 0.80%, respectively, and similar all-cause mortality of 1.7% and 2.0%, respectively. This implies an excellent safety profile by using these cut-offs. In addition, the study revealed a reduction in chest imaging (10%) and only an insignificant reversal of anticoagulation initiation (<1.5%). A prospective, multicentric diagnostic outcome trial comparing the YEARS algorithm to the standard cut-off in 3616 patients showed a reduction in imaging studies of 14% with a YEARS criteria adjusted D-dimer cut-off, while also confirming excellent safety (3-month missed VTE rate 0.61% and VTE-related fatality rate 0.2%) [12]. Our results are also confirmed by a recently published individual patient data meta-analysis of 16 studies including 20,553 patients that evaluated safety and efficiency of PTP assessment in combination with different D-dimer cut-offs (patient- and non-adjusted) [22]. This study revealed a failure rate (predicted VTE at 3 months after the exclusion of VTE without imaging) of 0.36%–0.58% for the standard cut-off, depending on whether the Wells or Revised Geneva Score was used, whereas it was 0.76%–1.1% for age-adjustment, and 1.8%–2.8% for YEARS adjustment. As implied by our investigation, efficiency (correct rule-out of VTE without imaging test) also increased significantly with the use of age-adjusted or YEARS algorithms in this study [22]. Thus, our data suggest that patient-adjusted D-dimer cut-offs are potentially safe to rule out VTE while having the substantial potential to limit unnecessary radiation and contrast media exposure and reducing health care costs.
Although the use of patient-adjusted D-dimer cut-offs seems promising for the future, there are some caveats that need to be borne in mind. Our study shows that the test performance of patient-adjusted D-dimer cut-offs varies substantially among the different subgroups. Although AUC was excellent for age- and YEARS adjustment, it was good for PTP adjustment and only fair for COVID-19-adjusted cut-offs. Thus, there seems to be a considerable variation depending on the type of patient-adjustment performed as reported by Stals etal. [22]. As AUC was not improved by adjustment, discrimination capacity did not increase, and the improvement in specificity came at the cost of sensitivity. Moreover, prevalence significantly influences the failure rate of the applied patient-adjustment [19, 33]. Thus, adjusting for patient factors implies finding the optimal trade-off between efficiency and safety, that is, failure rate. However, interpretation of failure rate is not always straightforward [33]. A higher failure rate at 3 months in a population with a high risk for VTE might also be caused by a new VTE event and not necessarily a missed VTE at the primary assessment [22]. In addition, the failure rate of the alternative—namely, sending the patient for imaging studies—also rises with increasing PTP [34]. Moreover, lastly, as shown by a recent individual patient data meta-analysis [22], efficiency and safety of patient-adjusted D-dimer testing vary significantly with heterogeneity in the population under examination (demographic factors, i.e., age >80 years, or comorbidities such as cancer or history of VTE) [35]. Although studies in the current literature more or less follow a “one-size-fits-all” approach and most often include the general emergency department population in their investigations, studies in distinct subgroups (i.e., age >80 years) or studies that combine several patient-adjustments to account important covariates (e.g., as reported by Freund etal. [20]) are currently scarce [36]. Such investigations could enhance individualized patient-centered care in the future. Thus, patient-adjusted D-dimer cut-offs have different performances and varying robustness depending on subgroups or factor adjusted for, prevalence of VTE, and heterogeneity of the examined study population. This might explain why available studies offer different conclusions regarding the best adaptation strategy when it comes to a direct comparison within the same cohort, depending on whether safety or performance is addressed as a priority [15, 37-40].
Limitations
Several limitations of our investigation warrant discussion. The heterogeneity of included trials in this review and meta-analysis—reflected, for instance, by high variance of VTE prevalence—is a major limitation, even though average prevalence of around 8% seems adequate for the evaluation of test characteristics. Further, heterogeneity and potential bias arise due to dissimilarity in recruitment strategies, inconsistent eligibility criteria, D-dimer assays used, and inconsequent implementation of clinical decision rules across studies.
An additional source of bias lies in the mainly observational nature of the available literature and in the low number of studies available in certain subgroups (e.g., COVID-19). The inclusion of retrospective studies—contributing to >50% (rough estimation) of the total study population—is a source of potential bias (selection and detection) because a clear diagnostic standard is often missing in such studies—or the follow-up is absent/incomplete, and it often remains unclear why imaging tests were performed in the first place. Further, these investigations often do not explain whether the reported prevalence and incidence of VTE represent actual relevant thrombosis or potentially less relevant calf vein thrombosis/sub segmental PE. We tried to mitigate this issue by showing prospective and retrospective investigations separately (see online Supporting Information). Further, only the minority of current investigations report on the PTP (low, medium, high) in their investigated patient populations; most trials report “non-high” pretest prevalence. It was thus impossible to evaluate the impact of patient-adjusted D-dimer values according to PTP. As outlined in the discussion, PTP has substantial impact on test characteristics and safety and thus is of major importance when evaluating a diagnostic test. Hence, this is a source of bias. We tried to account for this by analyzing trials according to PTP where respective data were available (see online Supporting Information).
Moreover, another issue that warrants discussion is the significant colinearity that exists between analyzed subgroups (e.g., age and renal function). This is why multiple-factor adjustment of the D-dimer cut-off might have some merit in the future. Lastly, the analysis of further important subgroups (e.g., tumor patients, COPD) was not possible, as most of the included studies did not present respective data. Additionally, a further important clinical question that remains is for which patient groups adjusted D-dimer cut-offs might not be accurate. This issue should be addressed in future prospective studies.
Conclusion
This large systematic review and meta-analysis, including more than 140,000 patients, confirm that the adjustment of D-dimer cut-off values to patient-specific factors seems safe—however, with considerable variation according to the applied strategy. The use of patient-adjusted D-dimer cut-offs revealed a significant decrease in false positives and thus embodies a large potential for reduction in imaging studies when patient-adjusted D-dimer cut-offs are used. Although test performance was best for YEARS algorithm, it was excellent for age-adjusted D-dimer cut-offs, good for PTP, and not sufficiently safe for COVID-19-adjusted D-dimer cut-offs. Although overall robustness, safety, and efficiency were good, the performance varied considerably among the different adjustment strategies with a high degree of heterogeneity—a fact that needs to be kept in mind when choosing and using such strategies.
Author contributions
Data curation; formal analysis; writing—review and editing: Tobias Krebs. Formal analysis; methodology; writing—review and editing: Martin Müller.
Conflict of interest statement
CAP, ASM, TK, DH, and JLG report grants from Orion Pharma, Abbott Nutrition International, B. Braun Medical AG, CSEM AG, Edwards Lifesciences Services GmbH, Kenta Biotech Ltd, Maquet Critical Care AB, Omnicare Clinical Research AG, Nestle, Pierre Fabre Pharma AG, Pfizer, Bard Medica S.A., Abbott AG, Anandic Medical Systems, Pan Gas AG Healthcare, Bracco, Hamilton Medical AG, Fresenius Kabi, Getinge Group Maquet AG, Dräger AG, Teleflex Medical GmbH, Glaxo Smith Kline, Merck Sharp and Dohme AG, Eli Lilly and Company, Baxter, Astellas, Astra Zeneca, CSL Behring, Novartis, Covidien, and Nycomed outside the submitted work. The money was paid into departmental funds and no personal financial gain applied. All other authors have nothing to disclose.
References
