Background Receiver operating characteristic (ROC) analysis is a common method used in diagnostic and screening tests to define thresholds levels of a factor that discriminates between 2 levels of another factor. The purpose of this analysis was to use ROC analysis to determine the optimal accelerometer-measured physical activity (PA) thresholds for predicting selective cardiovascular disease (CVD) risk factors. Methods ROC was performed using data from Stanford Girls Health Enrichment Multisite Studies trial. PA was assessed for multiple days using accelerometers. CVD variables were overweight, elevated triglyceride, reduced HDL-C, hypertension, impaired fasting glucose, fasting insulin, and clustering of multiple CVD risk factors. Results A sample of 261 girls participated, of which 208 had complete CVD risk measures (mean ± SD age = 9.4 ± 0.9yrs, BMI = 20.7 ± 4.8kg/m2). An average of ≥11.1 minutes/day at ≥2,600 counts/min was the maximally sensitive and specific threshold for discriminating girls who were overweight, ≥16.6 minutes/day at ≥2,000 counts/min for hyperinsulinemia or with ≥2 CVD risk factors. The Area Under the Curve for overweight, hyperinsulinemia, and ≥2 CVD risk factors was of 0.66, 0.58, and 0.60, respectively. The sensitivity and specificity associated with overweight, hyperinsulinemia, and ≥2 CVD risk factors were 60.3% and 72.9%, 53.3% and 83.9%, 44.0% and 84.7%, respectively. Conclusion Empirically-derived thresholds of PA to optimally discriminate between girls with and without CVD risk were lower in this sample than generally recommended. This ROC approach should be repeated in other populations to determine optimal PA thresholds with clinical validity for research, surveillance and program evaluation.