Predictive Model for Pulmonary Embolism in Patients with Deep Vein Thrombosis

Abstract

Background: To develop and verify a risk predictive model/scoring system for pulmonary embolism (PE) among hospitalized patients with deep venous thrombosis of the lower extremities (LDVT).

Methods: 776 patients with LDVT were enrolled in a case-control study between January 2016 and June 2017 from the Vascular Surgery Department of Shanxi Dayi Hospital, China. They were randomly divided into development (543 patients, 70%) and validation (233 patients, 30%) databases. Based on the results of pulmonary computed tomography arteriography, patients were divided into 2 categories; those with PE were designated as the case group, whereas those without comprised the controls. A logistic regression model and scoring system for PE in patients with LDVT was established in the development database and verified in the validation database. Scoring system (Shanxi Dayi Hospital score [SDH score]) was tabulated as follows: right lower extremity or bilateral lower extremities, 1; surgery or immobilization, 1; malignant tumor, 1; history of venous thromboembolism (VTE), 2; D-dimer >1,000 ng/mL, 2; and unprovoked, 2. Calibration and discrimination of the model were assessed by the Hosmer-Lemeshow goodness of fit test and the area under the receiver operating characteristic curve (AUC). Wells score, the Revised Geneva score, and the SDH score for predictive value of PE by AUC in the validation database were compared.

Results: 776 patients with LDVT were divided into 2 risk categories based on the scores from the risk model as follows: PE unlikely (score <3) and PE likely (score ≥3). Sensitivity, specificity, and crude agreement of the SDH score in the development database were 76.39%, 55.89%, and 61.33%, respectively. In the validation database, the logistic regression model showed good calibration and discriminative power. The Hosmer-Lemeshow goodness of fit test P value was >0.05, and the AUC was 0.705 (95% CI: 0.634-0.776, P < 0.001). The SDH score also showed good discriminative power, and the AUC was 0.702 (95% CI: 0.631-0.774, P < 0.001). Sensitivity, specificity, and crude agreement of the SDH score in the validation database were 67.61%, 61.73%, and 63.52%, respectively. AUC for the Wells score and the Revised Geneva score was 0.611 (95% CI: 0.533-0.688, P = 0.007) and 0.585 (95% CI: 0.503-0.666, P = 0.040), respectively. Difference of the AUC was not statistically significant between the Wells score and the SDH score (0.611 vs. 0.702, P = 0.059) but was so between the Revised Geneva score and the SDH score (0.585 vs. 0.702, P = 0.016). Sensitivity of the Wells score, Revised Geneva score, and the SDH score (64.79%, 67.61% vs. 67.61%) was not statistically significant. However, the specificity of the Wells score and Revised Geneva score was significantly lower than that of the SDH score (48.77%, 39.51% vs. 61.73%).

Conclusions: Our logistic regression model and the SDH score based on 7 risk factors as right lower extremity, bilateral lower extremities, unprovoked, surgery or immobilization, malignant tumor, history of VTE, and D-dimer>1,000 ng/mL showed good calibration and discriminative power for the assessment of PE risk in patients with LDVT. The SDH score is more specific for PE prediction in the Chinese population, compared with the Wells score and the Revised Geneva score.