Supervised machine learning on electrocardiography features to classify sleep in noncritically ill children.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Pediatric Sleep Medicine

Background:

• Real-time bedside sleep monitoring is unavailable in pediatric intensive care units, despite frequent sleep disruption.
• Cardiovascular dynamics are influenced by the autonomic nervous system during sleep, suggesting ECG as a potential data source.

Purpose of the Study:

• To develop and validate supervised machine learning models for automated sleep classification using electrocardiography (ECG) data in children.
• To assess the performance of these models across different age groups and sleep state classifications.

Main Methods:

• Retrospective analysis of polysomnography recordings from 90 noncritically ill children (6 months to 18 years).
• Extraction of ECG features in time, frequency, and nonlinear domains.
• Development of 2- to 5-state sleep classification models using logistic regression, random forest, and XGBoost, with 5-fold nested cross-validation.

Main Results:

• Models achieved an area under the receiver operator characteristic curve of 0.72-0.78.
• Balanced accuracies ranged from 0.70-0.72 for 2-state to 0.41-0.42 for 5-state classification.
• XGBoost generally performed best, except for 5-state classification where logistic regression excelled.

Conclusions:

• ECG-based machine learning models provide a promising, noninvasive method for bedside sleep classification in noncritically ill children.
• The models demonstrated moderate-to-good performance for 2- and 3-state sleep classification in this pediatric population.