This paper focuses on the REM sleep as one of sleep stages and proposes its estimation method based on the accelerometer attached to a nightwear around waist. In our approach, the REM sleep is estimated by (i) excluding other sleep stages before the REM sleep estimation and (ii) correcting the estimated probability according to the difference of the two scale moving averages. To investigate the effectiveness of the proposed method, this paper conducted the human subject experiment on 35 whole nights, which revealed the following implications: (i) the proposed method can significantly improve recall and F1 compared to Random Forest as one of the major machine learning; (ii) the REM sleep estimation excluding other sleep stages can improve performance compared to REM estimation without.Clinical relevance— A technology for estimating sleep on a daily basis is needed. The results show that the proposed method improves the performance of the sleep stage estimation by correcting inaccurate probabilities of the sleep stage. The proposed method can contribute to improving the performance of sleep stage estimation in daily life.