SnoozeNet: Causality-Driven Transformers for Interpretable Sleep Stage Classification

Sleep stage classification (SSC) is a critical procedure in sleep clinics, aiding in the diagnosis of various disorders such as sleep apnea and insomnia. However, automating this process using deep learning models often encounters the “black box” problem, which limits clinical adoption due to a lack of transparency. To address this issue, we developed SnoozeNet, a lightweight causality-based transformer model that enhances both interpretability and efficiency. Our interpretable mechanism is integrated through Granger causality, allowing our model to identify the most crucial features in each EEG data epoch. SnoozeNet demonstrates performance comparable to other state-of-the-art SSC technologies, all while setting a new standard for clinical applicability by prioritizing explainability and efficiency.