Consumer activity trackers such as Fitbit have been increasingly used in longitudinal studies to track sleep patterns. However, these devices are known to be inaccurate especially for measuring sleep stages. In this study we propose a two-stage classification method to predict more accurate sleep stage data from Fitbit data. Support vector machine was used in stage-1 classification to predict whether a Fitbit measurement is correct, and XGBoost algorithm was used in stage-2 classification to correct predicted wrong measurements. The results showed that our method improved Cohen’s Kappa and Matthews correlation coefficient by up to 14.3% and 13.3% respectively.