This article, presented by Le et al., introduces the SynthSleepNet as a multimodal hybrid-self-supervised learning frame work to evaluate polysomnography (PSG) data to better monitor the quality of sleep and diagnose various sleep disorders. This framework addresses the limitations of current methods that rely on large-scale-labeled datasets via combining masked prediction and contrastive learning to optimize complementary features from various multimodal physiological signals (EEG, EMG, etc.). To compare its performance against modern methods, a Mamba-based temporal context module (TCM) was created to gather information across signals. SynthSleepNet was found to have performed significantly better than these other methods with high accuracies across three tasks: the classification of sleep stages, detection of apnea, and detection of hypopnea. Respectively, for each tasks, the accuracies were determined to be 89.89%, 99.75%, and 89.60%. In a semi-hybrid-self-supervised learning environment with only 1% or 5% of the labeled dataset, this model was found to have maintained high accuracies up to 87.98%, 99.37%, and 77.52% in the three respective tasks as well. These findings highlight SynthSleepNet’s capability to efficiently and accurately analyze sleep-related physiological signals, even under data-constrained conditions.