ModelS4Apnea: leveraging structured state space models for efficient sleep apnea detection from ECG signals

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Sleep apnea is a prevalent disorder with serious health implications.
• Accurate and efficient detection methods are crucial for managing sleep apnea.
• Electrocardiogram (ECG) signals offer a non-invasive source for sleep apnea detection.

Purpose of the Study:

• To introduce ModelS4Apnea, a novel deep learning framework for automated sleep apnea detection.
• To leverage structured state space models (S4) for enhanced temporal modeling in ECG analysis.
• To achieve high accuracy and efficiency in sleep apnea detection from ECG spectrograms.

Main Methods:

• Developed a deep learning framework (ModelS4Apnea) integrating CNNs and S4 models.
• Utilized ECG spectrograms as input for the deep learning model.
• Trained and evaluated the model on the Apnea-ECG dataset.

Main Results:

• Achieved high performance metrics: 0.933 accuracy, 0.912 F1-score, 0.916 sensitivity, 0.944 specificity.
• Demonstrated superior performance compared to existing methods on the Apnea-ECG dataset.
• Showcased computational efficiency with fewer parameters and reduced training time.

Conclusions:

• ModelS4Apnea provides a robust and efficient solution for sleep apnea detection.
• The framework's low memory footprint and fast inference are suitable for real-time applications and wearables.
• The model's ability to perform per-recording classification highlights its diagnostic potential.