Can clinical features distinguish between immobile patients with stroke at high and low risk of deep vein thrombosis? Statistical modelling based on the CLOTS trials cohorts

BACKGROUND: Deep vein thrombosis (DVT) is an important complication of stroke. Guidelines recommend that DVT prophylaxis should be guided by an assessment of the individual patient's risk. The authors aimed to develop and test models to predict DVT risk. METHODS: The Clots in Legs Or sTockings after Stroke (CLOTS) Trial 1 randomised 2518 immobile patients with acute stroke to thigh-length graduated compression stockings (GCS) or no GCS and CLOTS Trial 2 randomised 3114 to thigh-length or below-knee GCS. The authors collected potential predictive variables at baseline and detected DVTs with compression duplex ultrasound scans at about 7-10 days and 25-30 days. The authors developed models with logistic regression to predict DVT in 1242 Trial 2 patients who had two scans and tested the models in the 1422 Trial 1 patients with two scans by estimating the area under the receiver operating characteristic curve (AUC). RESULTS: 168 (11.8%) patients in Trial 1 and 122 (9.8%) in Trial 2 had proximal DVTs. A model based on the Trial 2 cohort contained four of the 12 baseline variables: dependent before stroke (OR=3.62, 95% CI 2.15 to 6.08), unable to lift arms off bed (OR 1.89, 95% CI 1.23 to 2.90), history of DVT/pulmonary embolism (OR 3.69, 95% CI 1.98 to 6.88) and diabetes (OR 0.55, 95% CI 0.30 to 0.99). The AUC in the development cohort was 0.65 but only 0.57, 95% CI (0.53 to 0.61) in the Trial 1 cohort, indicating poor discrimination. CONCLUSIONS: Unfortunately, models based on clinical factors alone discriminate poorly between immobile patients with stroke at high and low risk, and would not facilitate individual tailoring of DVT prophylaxis strategies.