SViT: A Spectral Vision Transformer for the Detection of REM Sleep Behavior Disorder

Area of Science:

• Neurology
• Artificial Intelligence
• Sleep Medicine

Background:

• REM sleep behavior disorder (RBD) is a parasomnia characterized by dream enactment and REM sleep without atonia (RSWA).
• Manual polysomnography (PSG) scoring for RBD is time-consuming and subjective.
• Isolated RBD (iRBD) is a significant predictor of future Parkinson's disease.

Purpose of the Study:

• To introduce and evaluate a novel spectral vision transformer (SViT) for automated RBD detection using PSG signals.
• To compare the SViT's performance against conventional convolutional neural network (CNN) architectures.
• To investigate the diagnostic utility of different PSG channels (EEG, EMG, EOG) for RBD detection using deep learning.

Main Methods:

• Deep learning models, including SViT and CNN, were applied to scalograms derived from PSG data (EEG, EMG, EOG).
• A 5-fold bagged ensemble approach was utilized with 153 RBD patients (96 iRBD, 57 RBD with PD) and 190 controls.
• Model performance was assessed per-epoch and per-patient, with SViT interpretation using integrated gradients.

Main Results:

• The SViT achieved the best per-patient performance with an F1 score of 0.87.
• When trained on channel subsets, the SViT reached an F1 score of 0.93 using EEG and EOG data.
• Model interpretation highlighted the diagnostic relevance of EEG and EOG, challenging the traditional emphasis on EMG.

Conclusions:

• The SViT demonstrates superior per-patient diagnostic accuracy for RBD compared to CNNs.
• EEG and EOG channels are crucial for deep learning-based RBD detection, potentially streamlining the diagnostic process.
• This AI-driven approach offers a promising avenue for efficient and accurate diagnosis of RBD and early identification of neurodegenerative disease risk.