Explainable vision transformer for automatic visual sleep staging on multimodal PSG signals.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Sleep Medicine

Background:

• Polysomnography (PSG) scoring is vital for sleep disorder diagnosis but is subjective and time-consuming.
• Existing machine-learning models for PSG lack clinical interpretability due to their 'black-box' nature.

Purpose of the Study:

• To develop an interpretable automatic sleep staging system, SleepXViT, using Vision Transformer (ViT).
• To enhance the reliability and clinical utility of AI in sleep scoring.

Main Methods:

• SleepXViT employs a Vision Transformer (ViT) architecture to mimic human visual scoring of PSG data.
• The system was evaluated on the KISS-a dataset (7745 patients) and public datasets (SHHS1, SHHS2).
• Interpretability features include confidence scores, heatmaps, and relevance scores.

Main Results:

• SleepXViT achieved a Macro F1 score of 81.94% on the KISS-a dataset, outperforming baseline models.
• The system demonstrated robust performance on public sleep datasets.
• Explanations like heatmaps and confidence scores were generated to enhance interpretability.

Conclusions:

• SleepXViT provides reliable and interpretable automatic sleep staging, addressing the limitations of traditional PSG scoring and black-box AI models.
• The system's explainability features facilitate synergy between AI and human scorers in clinical settings.
• SleepXViT shows promise for improving the efficiency and consistency of sleep disorder diagnosis.