Daytime sleepiness is not only the cause of productivity decline and accidents, but also an important metric of health risks. Despite its importance, the long-term quantitative analysis of sleepiness in daily living has hardly been done due to time and effort required for the continuous tracking of sleepiness. Although a number of sleepiness detection technologies have been proposed, most of them focused only on driver’s drowsiness. In this paper, we present the first step towards the continuous sleepiness tracking in daily living situations. We explore a methodology for predicting subjective sleepiness levels utilizing respiration and acceleration data obtained from a novel wearable sensor. A class imbalance handling technique and hidden Markov model are combined with supervised classifiers to overcome the difficulties in learning from an imbalanced and time series dataset. We evaluate the performance of our models through a comprehensive experiment.