Background: Arousals during sleep become pathological when the frequency of occurrence increases beyond the normal limit. The visual scoring of arousals routinely conducted by sleep experts is a challenging task warranting an automatic approach. This paper presents a framework for automatic, multimodal detection of arousals during sleep using subject-specific and subject-independent modeling. Method: The training and test sets for the challenge include physiological signals (EEG, EOG, EMG, EKG, and SaO2) for 994 and 989 subjects, respectively. The training set was split to 796 and 198 subjects as in-home train and test subjects. Two modeling approaches were tested: 1) subject-specific: a separate model for every training subject, and 2) subject-independent: one model for all the training subjects. The resulting models were tested on in-home test subjects. In the subject-specific model, a feature-based approach was implemented using band-specific EEG energies and SaO2 variability features computed for non-overlapping 5-minute windows as features and a support vector machine as a classifier. For the subject-independent model, an ensemble of a feature-based approach (similar to the subject-specific model) and a convolutional neural network (CNN) was used. The CNN was trained on spectrogram of non-overlapping 30-second EEG segments. Results: The subject-specific model achieves an area under curve (AUC) of 0.596 on the challenge test set (AUprc= 0.063). The results achieved at the unofficial phase of the challenge on the in-home test set for subject-specific and subject-independent were AUC of 0.611 and 0.564, respectively. Discussion: In this study, subject-specific and subject-independent models were tested for automatic detection of arousals during sleep. The initial performance of the proposed models encourages us to further improve the algorithms in the official phase of the challenge using additional features especially from other modalities as well as creating new CNN models for other physiological signals.