Risk prediction scores have been devised to identify patients at increased risk for Venous Thromboembolism (VTE) in different patient populations and settings. Guideline recommendations for VTE risk assessment vary greatly. We performed a systematic review to synthesize evidence on clinical risk prediction scores for VTE in hospitalized medical and surgical patients.

We systematically searched Medline, EMBASE, Cochrane, National Institute of Health and Care Excellence (NICE), National Guidelines Clearinghouse (NGC), and Guidelines International Network (GIN) databases up to March 2016. We included studies validating risk prediction scores for adult hospitalized patients. We excluded studies for any of the following reasons: non-English publication, conducted in non-OECD (Organisation for Economic Co-operation and Development) countries, validation cohorts focused solely on critical care patients, or scores developed for specific surgical or medical sub-specialty populations. We plotted receiver operating characteristic (ROC) curves of included studies and performed summary ROC meta-analyses for scores in which >1 external validation studies were combinable. Risk of bias was assessed qualitatively. We assessed the strength of the evidence base using Grading of Recommendations Assessment, Development and Evaluation (GRADE).

We screened 110 primary studies and included 18 of those for analysis. There were seven studies of the Caprini score, three studies of the Padua score, two studies of the IMPROVE score; and one study each of the Arcelus, Geneva, Khorana, RAP, and Kucher scores . Strength of evidence was downgraded for study risk of bias because most studies disproportionately included patients at high risk of VTE. Our summary estimates of the performance of the three combinable scores at clinically-relevant thresholds are: Caprini score at a threshold of three in surgical patients – 96 percent sensitivity, 44 percent specificity; IMPROVE at a threshold of one in medical patients – 96 percent sensitivity, 20 percent specificity; and Padua at a threshold of 4–87 percent sensitivity and 58 percent specificity.

There is moderate strength evidence for use of the Caprini score to predict VTE in surgical patients and for the Padua and IMPROVE scores in medical patients. Lower thresholds may be warranted to achieve sufficient sensitivity to identify low risk populations who may not require routine VTE prophylaxis. Studies making direct comparisons of risk prediction scores in similar patient populations are lacking and are necessary to ascertain which score is most effective.