<p>Monitoring and identifying sleep postures, states or stages is a prerequisite for diagnosing sleep disorders, and provides crucial information for the treatment of sleep disorders. At present, there is no universal multimodal sleep recognition model, and most of the existing methods are limited to accurate modal sleep data. To address these challenges, we propose a dynamic saliency enhanced fusion network, which ranks the features based on data saliency and dynamically selects feature vectors with high attention and coherence, a two-block module is designed for multi-scale convolution and sequence features. We utilized datasets from three distinct sources (the Sleep Bioradiolocation Dataset, the Pressure Map Dataset, and the PSG Dataset), dataset-collecting over 100,000 training samples from dozens of healthy subjects and patients with Parkinson’s disease for our model to learn from. The experimental results show that the model has a 2% -3% improvement compared to existing algorithms, with a promising practical application prospects in sleep recognition applications such as smart healthcare and smart home.</p>