Lateral connection convolutional neural networks for obstructive sleep apnea hypopnea classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Sleep Medicine

Background:

• Convolutional Neural Networks (CNNs) are effective for obstructive sleep apnea hypopnea (OSAHS) classification but lack interpretability.
• Current models require extensive labeled data, which is costly and time-consuming to acquire.
• Lateral connections are vital in visual neurobiology but underexplored in CNNs for medical applications.

Purpose of the Study:

• To introduce a novel CNN architecture, the Lateral Connection CNN (LCCNN), for improved OSAHS classification.
• To enhance model interpretability and reduce reliance on labeled datasets through unsupervised learning.
• To investigate the impact of integrating lateral connections into CNNs for sleep apnea diagnosis.

Main Methods:

• Developed a novel LCCNN architecture with convolution, lateral connection, competition, and pooling layers.
• The competition layer enables unsupervised filter updates and semantic neuron arrangement.
• Evaluated LCCNN performance on the University College Dublin (UCD) and Physionet Challenge (PCD) databases.

Main Results:

• LCCNN achieved high classification accuracies: 97.3% (kappa 0.9) on UCD and 95.6% (kappa 0.83) on PCD.
• Demonstrated effective feature extraction and salient wave detection through lateral connections.
• The unsupervised competition layer facilitated semantic arrangement of neurons.

Conclusions:

• The LCCNN model offers a promising, interpretable, and data-efficient approach for OSAHS classification.
• This architecture provides a foundation for future research in unsupervised deep learning for sleep disorder diagnosis.
• Integrating lateral connections into CNNs can significantly improve performance in medical image analysis.