Adaptive Margin Aware Complement-Cross Entropy Loss for Improving Class Imbalance in Multi-View Sleep Staging Based...

Area of Science:

• Machine Learning
• Sleep Medicine
• Computational Neuroscience

Background:

• Sleep data often exhibits class imbalance, biasing models towards common sleep stages and reducing accuracy for clinically significant minority stages.
• Accurate classification of minority sleep stages (e.g., N1) is vital for diagnosing disorders like hypersomnia and narcolepsy.

Purpose of the Study:

• To develop an improved sleep staging classification model that addresses class imbalance issues.
• To enhance the accuracy of minority sleep stage classification using a novel loss function.

Main Methods:

• Proposed a multi-view Convolutional Neural Network (CNN) model incorporating an adaptive margin-aware loss function.
• Introduced a novel margin-aware factor considering relative sample sizes to increase regularization for minority classes.
• Developed margin-aware cross-entropy and margin-aware complement entropy loss functions, with the latter neutralizing errors for minority classes.

Main Results:

• The proposed adaptive margin-aware loss function significantly improved overall sleep staging accuracy across three datasets.
• Demonstrated substantial improvements in the classification accuracy of minority sleep stages compared to state-of-the-art methods.
• The margin-aware complement entropy effectively increased regularization and neutralized errors for minority classes.

Conclusions:

• The novel adaptive margin-aware loss function effectively mitigates class imbalance in sleep staging.
• The proposed method offers a promising approach for improving the diagnosis of sleep disorders through enhanced sleep stage classification.
• This technique provides a significant advancement in accurately identifying clinically relevant minority sleep stages.