A Predictive Model To Identify Patients with Venous Thromboembolism (VTE) at Minimal Risk of Malignancy.

Association between VTE and cancer has been recognised for over a century. The incidence of occult or overt malignancy in patients with thrombosis is 7–26%. VTE has been shown to impart poor prognosis in patients with cancer and cancer also adversely impacts on the survival of patients with VTE. It is desirable to identify patients with thrombosis at increased risk of malignancy or those with minimal risk, thus avoiding unnecessary investigations. We propose here a predictive model using age and quantitative D-dimer level at presentation along with site of thrombosis.

Materials and Methods

This study included 696 (M: 358; F: 338) consecutive patients from the prospectively maintained database of patients with venous thrombosis at a University Teaching Hospital, between February 2001 and December 2005. All patients underwent a Doppler ultrasound examination to confirm the diagnosis and determine the extent of venous thrombosis. At presentation, D-dimer assays were done using Bio-Merieux kit containing mouse monoclonal antibody. The database was regularly updated (6 monthly) using hospital information systems, questionnaires and clinical review Thrombosis recurrence was always confirmed by Doppler ultrasound examination. All Patients with thrombosis received standard treatment with low molecular weight heparin and warfarin. Statistical analysis was carried out using SPSS 13.0 for Windows software. A logistic multivariate regression model was fitted using complete data records from 621 patients with an indicator variable for a subsequent cancer as the response and age, the natural logarithm of the quantitative D-dimer level and the site of the thrombosis as explanantory variables. The fitted model was validated using an additional set of independent data.

Results

The model correctly identified the VTE patients without malignancy in 473 out of 480 cases (98.5% accuracy). But the model was ineffective in identifying VTE patients with malignancy (13 out of 141; 9% accuracy). The area under the receiver operating characteristic (ROC) curve was 0.72, indicating that the test developed for predicting cancer for the model data was reasonably good. Our model shows that below a predicted probability of 0.10 less than 5% of the patients actually developed cancer (9/190) whereas for a predicted probability of 0.19 less than 10% of patients actually developed cancer (27/276). In the validation dataset of 93 patients with VTE, the model correctly identified the number of patients without malignancy in 72 out of 73 cases (98.6% accuracy). One-sample binomial tests revealed that there was no significant difference in the number VTE patients with cancer between the model and the validation dataset for the predicted probabilities of 0.10 and 0.19 (p-values of 0.650 and 0.246 respectively).

Conclusions

This suggests that our model is useful for identifying patients at minimal risk of developing cancer with VTE. This predictive model has been validated by an independent dataset demonstrating reproducibility. This model will enable a focused and a cost-effective strategy of screening for a malignancy in patients with VTE.