External validation of the Vascular Study Group of New England carotid endarterectomy risk predictive model using an independent U.S. national surgical database.

OBJECTIVE: Previously, we described a Vascular Study Group of New England (VSGNE) risk predictive model to predict composite adverse outcomes (postoperative death, stroke, myocardial infarction, or discharge to extended care facilities) after carotid endarterectomy (CEA). The goal of this study was to externally validate this model using an independent database. METHODS: The American College of Surgeons National Surgical Quality Improvement Program (NSQIP) CEA-targeted database (2010-2014) was used to externally validate the risk predictor model of adverse outcomes after CEA previously created using the VSGNE carotid database. Emergent cases and those in which CEA was combined with another operation were excluded. Cases in which a discharge destination cannot be determined were also excluded. To assess the predictive power of our VSGNE prediction score within this sample, a receiver operating characteristic curve was constructed. Risk scores for each NSQIP patient were also computed using beta weights from the VSGNE CEA model. To further assess the construct validity of our VSGNE prediction score, the observed proportion of adverse outcomes was examined at each level of our prediction scale and within five roughly equally sized risk groups formed on the basis of our VSGNE prediction scores. RESULTS: In this database, 10,889 cases met our inclusion criteria and were used in this analysis. The overall rate of adverse outcomes in this cohort was 8.5%. External validation of the VSGNE model on this sample showed moderately good predictive ability (area under the curve = 0.745). Patients in progressively higher risk groups, based on their VSGNE model scores, exhibited progressively higher rates of observed adverse outcomes, as predicted. CONCLUSIONS: The VSGNE CEA risk predictive model was externally validated on an NSQIP CEA-targeted sample and showed a fairly accurate global predictive ability for adverse outcomes after CEA. Although this model has a good population concordance, the lack of cut point indicates that individual risk prediction requires more evaluation. Further studies should be geared toward identification of variables that make this risk predictive model more robust.