Aims: The study aims to provide a comprehensive classification method based on the nature of these arousals, incorporating a hidden markov model(HMM) and random decision forests(RF). Methods: For feature selection, windows of 20000 samples(10 seconds) were obtained from the 20% of the given training data. These windows were chosen so that their median was aligned with the starting point of the arousals. Then windows of the same size were randomly selected from regions that contain no arousal. 222 features were assessed using all the given channels, then reduced to 32 using attribute selection methods. These were either standalone features from the arousal part of the windowed data or the ratio of two consecutive smaller windows. The performance of the features were rated using RF and the area under the receiver operating characteristic curve(AUC-ROC). Part of the training data was processed again to obtain a sequence of stages based on arousal data. These stages include the start and the end of arousals. We applied a HMM trained on the annotated data, and changed the probability model of its states to random decision forests. Viterbi algorithm was used to predict arousals. Results: Beginning of arousals predicted only by RF using 10 fold cross-correlation received a 0.822 AUC-ROC on training data. RF with HMM received the same score for the start, a 0.841 for the end of the arousal and a 0.76 for the middle. For selecting the arousal intervals given by the Viterbi path, the detected intervals were scored by the probabilities got from random decision forests. Conclusion: The addition of HMM to RF improves the model as it enforces our prediction to follow the sequential nature of arousal and non-arousal sleep periods.