Sleep versus wake classification from heart rate variability using computational intelligence: consideration of...

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Pediatric Sleep Medicine

Background:

• Reliable classification is crucial for biomedical applications, especially with large datasets.
• Manual sleep state scoring from polysomnography (PSG) is time-consuming and labor-intensive.
• Developing automated, reliable methods for sleep analysis is highly desirable.

Purpose of the Study:

• To develop and validate a technique for reliable sleep/wake determination in infants using solely electrocardiogram (ECG) data.
• To assess the performance of machine learning classifiers (LVQ, MLP, SVM) in accurately classifying sleep states from ECG.
• To evaluate the impact of rejecting unreliable data segments on classification accuracy and efficiency.

Main Methods:

• Utilized 8-hour simultaneous ECG and PSG recordings from 190 infants in the CHIME study.
• Trained and tested Learning Vector Quantization (LVQ), Multilayer Perceptron (MLP), and Support Vector Machines (SVM) classifiers.
• Implemented a systematic rejection strategy for segments classified with low reliability.

Main Results:

• Classifiers achieved 85%-87% correct classification accuracy after rejecting approximately 30% of the data.
• The Kappa statistic ranged from 0.65-0.68, indicating substantial agreement.
• Accuracy improved by approximately 8% compared to models without rejection, demonstrating the benefit of reliability-focused classification.

Conclusions:

• A reliable sleep/wake classifier based solely on ECG is feasible, offering high accuracy, simplicity, and low intrusiveness.
• Integrating reliability assessment and rejection directly into the classification model enhances performance for infant sleep analysis.
• This ECG-only approach has significant potential for non-intrusive, large-scale infant sleep monitoring.