SleepBoost: a multi-level tree-based ensemble model for automatic sleep stage classification.

Area of Science:

• Neuroscience and Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Sleep disruption is strongly linked to neurodegenerative diseases, necessitating accurate sleep stage monitoring.
• Deep learning (DL) models have advanced automatic sleep stage classification (ASSC), but their opacity hinders clinical trust.
• Existing ASSC methods lack transparency, posing challenges for medical practitioners.

Purpose of the Study:

• To introduce SleepBoost, a transparent, multi-level tree-based ensemble model for ASSC.
• To enhance clinical trust and adoption of ASSC models by prioritizing interpretability.
• To provide a robust and accessible tool for sleep pattern analysis in neurodegenerative disorders.

Main Methods:

• Developed SleepBoost, a transparent multi-level tree-based ensemble model for ASSC.
• Implemented a feature engineering block (FEB) extracting 41 time/frequency domain features, with 23 selected via mutual information.
• Integrated three linear models within a multi-level tree structure, using adaptive weight allocation.

Main Results:

• SleepBoost achieved 86.3% accuracy, 80.9% F1-score, and 0.807 Cohen kappa on the Sleep-EDF-20 dataset.
• The model outperformed leading deep learning models in ASSC.
• Ablation studies confirmed the importance of selective feature extraction for accuracy and interpretability.

Conclusions:

• SleepBoost offers a transparent and high-performing alternative to opaque deep learning models for ASSC.
• The model's interpretability is crucial for clinical adoption in monitoring neurodegenerative diseases.
• Open-source availability promotes accessibility and potential widespread use in clinical settings.