A Deep Generative Model for Five-Class Sleep Staging with Arbitrary Sensor Input | JoVE Visualize

Area of Science:

Computational neuroscience
Machine learning for healthcare
Sleep medicine

Background:

Current sleep scoring relies on electroencephalography (EEG), electrooculography (EOG), and electromyography (EMG) signals.
Polysomnography includes numerous other signals (e.g., cardio-respiratory) that are underutilized for sleep staging.
Utilizing diverse signals can enhance reliability, resilience to signal loss, and enable long-term, non-obtrusive recordings.

Purpose of the Study:

To develop a deep generative model for automatic sleep staging using a flexible combination of multiple sensor modalities.
To achieve zero-shot inference on arbitrary sensor sets by leveraging a novel Bayesian factorization.
To introduce a new interpretability metric for evaluating sensor contributions to classification performance.

Main Methods:

Developed a score-based diffusion model trained on 1947 expert-labeled overnight polysomnographic recordings with 36 different signals.
Employed a novel Bayesian factorization of the score function for zero-shot inference across varying sensor combinations.
Proposed and validated an information gain-based interpretability metric correlated with classification performance.

Main Results:

The model achieved performance comparable to the polysomnography inter-rater agreement limit on single-channel EEG (5-class accuracy 85.6%, Cohen's kappa 0.791).
Demonstrated flexibility with unconventional sensor sets, achieving 79.0% accuracy (kappa 0.697) using finger photoplethysmography, nasal flow, and thoracic movements.
Showcased adaptability with tibialis and sternocleidomastoid EMG (71.0% accuracy, kappa 0.575), and the ability to incorporate new sensors post-hoc.

Conclusions:

The developed deep generative model offers a flexible and robust approach to automatic sleep staging across diverse sensor configurations.
The model achieves high performance, comparable to gold standards, and maintains accuracy with limited or unconventional sensor inputs.
The proposed interpretability metric and post-hoc sensor addition capability enhance the model's practical utility and adaptability in sleep research.