|Jago, Russell - UNIVERISITY OF BRISTOL|
Submitted to: Journal of Sports Sciences
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: July 4, 2006
Publication Date: June 1, 2007
Citation: Jago, R., Zakeri, I., Baranowski, T., Watson, K. 2007. Decision boundaries and receiver operating characteristic curves: New methods for determining accelerometer cutpoints. Journal of Sports Sciences. 25(8):937-944. Interpretive Summary: There has been concern about what cut points should be used with minute-by-minute counts using accelerometers to separate light from moderate physical activity. This manuscript reports a study where children wore accelerometers and were instructed in doing light activity and moderate activity. A method called the receiver operating characteristic curve was then used to identify the best cut point separating these two levels of physical activity. This method should be used in larger data sets.
Technical Abstract: We propose and evaluate the utility of an alternative method (decision boundaries) for establishing physical activity intensity-related accelerometer cutpoints. Accelerometer data collected from 76 11- to 14-year-old boys during controlled bouts of moderate- and vigorous-intensity field physical activities were assessed. Mean values and standard deviations for moderate- and vigorous-intensity activities were obtained and normal equivalents generated. The decision boundary (the point of intersection of overlapping distributions) was used to create a lower-bound vigorous-intensity cutpoint. Receiver operating characteristic (ROC) curves compared the sensitivity and specificity of the new cutpoint and mean values with the actual activity. There was a 96.5% probability that participants performing vigorous-intensity physical activity were accurately classified when using the decision boundary of 6700 counts per minute, in contrast to the 50% accurately classified when the mean value was used. Inspection of the empirical ROC curve indicated that the decision boundary provided the optimal threshold to distinguish between moderate and vigorous physical activity for this dataset. In conclusion, decision boundaries reduced the error associated with determining accelerometer threshold values. Applying these methods to accelerometer data collected in specific populations will improve the precision with which accelerometer thresholds can be identified.