MultimodalSleepNet: A Lightweight Neural Network Model for Sleep Staging based on Multimodal Physiological Signals

With social development, the demand for automatic sleep quality assessment in wearable devices is increasing, especially as sleep quality is closely related to the diagnosis of psychiatric disorders. However, existing automatic sleep stage classification models are mostly designed for unimodal signals, with an emphasis on increasing parameter scale and model depth. As a result, it is challenging to meet the requirements for both lightweight models and high accuracy in wearable device-based automatic sleep staging tasks. To address this problem, this study introduces a novel lightweight model for sleep staging, MultimodalSleepNet, which is based on multi-modal physiological signals. Specifically, the model is designed to capture the temporal dynamics of physiological signals and the spatial interactions between multimodal signals. Additionally, an inflationary convolution mechanism is incorporated to accelerate temporal feature extraction. We validate the model using the publicly available Sleep-EDF-Expanded dataset. Compared to similar studies, our model achieves outstanding performance, with accuracies of 93.1% and 90.2% in the three-stage and five-stage sleep recognition tasks, respectively. Notably, the three-stage classification results show an 11.9% improvement in modal fusion accuracy compared to unimodal signals, with an 8.9% improvement in multiclass F1 score and a 20.8% increase in Cohen's kappa coefficient. In conclusion, our study offers a reference for the design of lightweight models for sleep staging and provides new insights into feature extraction and fusion of multimodal signals.