STQS: Interpretable multi-modal Spatial-Temporal-seQuential model for automatic Sleep scoring.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Manual sleep scoring is time-consuming and subjective.
• Machine learning offers automated solutions but often lacks interpretability for clinicians.
• Deep learning models can potentially bridge this gap by providing accurate and understandable sleep scoring.

Purpose of the Study:

• To develop and evaluate a novel deep learning architecture (STQS) for multi-modal sleep scoring.
• To investigate the interpretability of the STQS model's decision-making process.
• To compare the model's reasoning against established American Academy of Sleep Medicine (AASM) guidelines.

Main Methods:

• Developed STQS architecture combining Convolutional Neural Networks (CNNs) for spatio-temporal feature extraction and Bidirectional Long Short-Term Memory (Bi-LSTM) for sequential information.
• Utilized three modalities: Electroencephalography (EEG), Electrooculography (EOG), and Electromyography (EMG).
• Incorporated residual connections to integrate features and evaluated model performance on two large datasets (SHHS and an in-house dataset).

Main Results:

• Achieved high accuracy (85% on SHHS, 77% on in-house data) and macro F1 scores (79% and 73%, respectively).
• Demonstrated that LSTM layers significantly improve performance over CNNs alone, while residual connections offered no additional benefit.
• Interpretability analysis confirmed the model's outputs align with AASM scoring guidelines, indicating domain knowledge consistency.

Conclusions:

• The STQS deep learning model provides accurate and interpretable multi-modal sleep scoring.
• The model's alignment with AASM guidelines enhances clinical trust and adoption potential.
• Future research should prioritize enhancing multi-modal approaches for sleep scoring over single-channel models.