Unlocking transcranial FUS-EEG feature fusion for non-invasive sleep staging in next-gen clinical applications

Abstract

Accurate and non-invasive sleep staging is essential for evaluating sleep quality and diagnosing neurological and sleep disorders. Addressing the variations in electroencephalogram (EEG) and electrooculogram (EOG) signals across different sleep stages, this study introduces a transcranial focused ultrasound (tFUS) based multimodal feature fusion deep learning model (MFDL) for automated sleep staging. The proposed framework integrates two one-dimensional convolutional neural networks (1D-CNNs) to extract sleep-relevant features from EEG and EOG signals, followed by an adaptive feature fusion module that dynamically assigns weights based on feature significance. By enhancing discriminative features and suppressing irrelevant ones, the model generates a robust multimodal representation of sleep information. Furthermore, a bidirectional long short-term memory (Bi-LSTM) network captures temporal dependencies in sleep stage transitions, improving classification accuracy. The effectiveness of MFDL is validated on the publicly available Sleep-EDF dataset, achieving 94.1% accuracy, 88.2% Kappa coefficient, and 81.9% MF1 score. Notably, the recall rates for the challenging N1 and REM sleep stages are significantly enhanced to 64.6% and 93.5%, respectively. These results highlight the potential of MFDL in enhancing tFUS-based neuromodulation by providing precise, data-driven sleep state monitoring, paving the way for advanced non-invasive brain stimulation technologies in next-gen clinical applications.