Performance of the Wells score in predicting deep vein thrombosis in medical and surgical hospitalized patients with or without thromboprophylaxis: The R-WITT study

Abstract

The Wells score had shown weak performance to determine pre-test probability of deep vein thrombosis (DVT) for inpatients. So, we evaluated the impact of thromboprophylaxis on the utility of the Wells score for risk stratification of inpatients with suspected DVT. This bicentric cross-sectional study from February 1, 2018 to January 31, 2019 included consecutive medical and surgical inpatients who underwent lower limb ultrasound study for suspected DVT. Wells score clinical predictors were assessed by both ordering and vascular physicians within 24 h after clinical suspicion of DVT. Primary outcome was the Wells score’s accuracy for pre-test risk stratification of suspected DVT, accounting for anticoagulation (AC) treatment (thromboprophylaxis for ⩾ 72 hours or long-term anticoagulation). We compared prevalence of proximal DVT among the low, moderate and high pre-test probability groups. The discrimination accuracy was defined as area under the receiver operating characteristics (ROC) curve. Of the 415 included patients, 30 (7.2%) had proximal DVT. Prevalence of proximal DVT was lower than expected in all pre-test probability groups. The prevalence in low, moderate and high pre-test probability groups was 0.0%, 3.1% and 8.2% ( p = 0.22) and 1.7%, 4.2% and 25.8% ( p < 0.001) for inpatients with or without AC, respectively. Area under ROC curves for discriminatory accuracy of the Wells score, for risk of proximal DVT with or without AC, was 0.72 and 0.88, respectively. The Wells score performed poorly for discrimination of risk for proximal DVT in hospitalized patients with AC but performed reasonably well among patients without AC; and showed low inter-rater reliability between physicians. ClinicalTrials.gov Identifier: NCT03784937.

Comparison of Two-Point Compression Ultrasonography Performed by Emergency Medicine Doctors and Radiologists in Detection of Deep Vein Thrombosis

“…Wells score is one of the best-known and most commonly used scoring systems in diagnosing DVT and pulmonary embolism. 12,13 According to the results of this study, the incidence of DVT in patients with a Wells score of 2 and below (low probability) was significantly lower than in those with a Wells score above two (21.8% vs. 47.1%). It is an expected result that the likelihood of DVT decreases as the Wells score drops, and it is compatible with previous studies on this subject.…”

In-Hospital versus Out-of-Hospital Pulmonary Embolism: Clinical Characteristics, Biochemical Markers and Echocardiographic Indices

“…When Söhne M, in a retrospective study, investigated the sensitivity, the negative predictive value of a non-high clinical probability of PE (based on Wells score), and a normal D-dimer result in outpatients and inpatients with suspected PE, they found that these tests were less reliable in inpatients compared to outpatients, especially in the elderly [32]. Also, the reliability of the scores for the discrimination of the risk for VTE in hospitalised patients differed among patients with or without anticoagulation treatment [33]. Furthermore, regarding the Geneva score and its revised form (revised Geneva score), it is known that both scores have been validated only in outpatients [8].…”

Machine Learning-Based Predictive Models for Patients with Venous Thromboembolism: A Systematic Review

“…unnecessary imaging, and reduce the risk of complications. Until now, pretest probability tests such as Well's score 22 and Geneva 23 based on clinical and vital signs have been validated in several studies and showed high negative predictive values but low positive predictive values 24 , moderate accuracy and a high dependence on the population studied 25 . iv) Personalized anticoagulant management: ML-CPMs can be used to predict the response to different anticoagulant therapies or identify a personalized dosage.…”

“…Until now, pretest probability tests such as Wells score 22 and Geneva 23 based on clinical and vital signs have been validated in several studies and showed high negative predictive values but low positive predictive values, 24 moderate accuracy, and a high dependence on the population studied. 25…”

“…For this reason and to be able to compare the predictive models we decided to focus on studies that processed specific group of data, by excluding NLP methods, image processing and computed tomography and narrowing down the research to models that have been developed directly from structured data extracted from the EHR, such as laboratory tests, genetic and clinical information. Only a few studies that used combined methods of data extraction (structured and unstructured) were included in our study[25][26][27] . Regarding genetic information, we only included studies derived from EHR data and not genomics, metabolomics, and proteomics data, since these do not represent typical information found in the current EHR system.…”