Machine learning classifier solving the problem of sleep stage imbalance between overnight sleep

Area of Science:

• Artificial Intelligence
• Machine Learning
• Sleep Medicine

Background:

• Manual sleep stage scoring is time-consuming, requires expertise, and is prone to subjective bias.
• Machine learning offers automated solutions but struggles with imbalanced datasets common in sleep studies.
• Poor performance on minority classes in sleep phase prediction hinders reliable automated analysis.

Purpose of the Study:

• To overcome data imbalance issues in machine learning for sleep scoring.
• To improve the generalization of sleep data for data-centric artificial intelligence.
• To evaluate methods for enhancing the accuracy of automated sleep phase prediction.

Main Methods:

• Applied feature extraction based on American Academy of Sleep Medicine (AASM) standards.
• Experimented with loss function adjustment and resampling techniques to address minority class prediction errors.
• Utilized various machine learning classifiers, adjusting datasets with sampling and class weighting.

Main Results:

• Achieved a best-performing model accuracy of 91.9% for sleep stage classification.
• Obtained a kappa score of 0.899 and an F1-score of 86.9% with the optimized model.
• Demonstrated the effectiveness of data balancing techniques in improving machine learning performance for sleep scoring.

Conclusions:

• Loss function adjustment and resampling effectively mitigate data imbalance in sleep scoring.
• Automated sleep phase prediction shows significant potential for clinical applications.
• Further research can explore channel accuracy and electrode monitoring for enhanced real-world performance.