Obstructive sleep apnea (OSA) is a common sleep disorder. Positive airway pressure (PAP) therapy is the first-line treatment, while its effectiveness is significantly limited by incomplete adherence in many patients. This work aims to find a predictive association between data from in-laboratory sleep studies during treatment (PAP titration polysomnogram, or PSG) and PAP adherence. Based on a PAP titration PSG database, we present a pipeline to develop a wavelet-based deep learning model and address two challenges. First, to tackle the problem of extremely long overnight PSG signals, it randomly draws segments and extracts features locally. The global representation for the entire signal is achieved by local feature P-norm pooling. Second, to tackle the problem of limited dataset size, the pre-trained EfficienNet-B7 is used as an unsupervised feature extractor to transfer ImageNet knowledge to PSG signals in the wavelet domain. The trained pipeline achieves 78% balanced accuracy and 83% AUC on the test set using airflow and frontal EEG signals, which, we believe, is a compelling result as a pilot study.