A multifaceted approach for obstructive sleep apnea classification from ECG signal using deep learning

Area of Science:

• Cardiology
• Sleep Medicine
• Biomedical Engineering

Background:

• Obstructive sleep apnea (OSA) presents significant cardiovascular and neurocognitive risks.
• Current diagnostic methods like polysomnography are expensive and impractical for widespread screening.
• Electrocardiogram (ECG) signals offer a viable, non-invasive, and cost-effective alternative for sleep apnea detection.

Purpose of the Study:

• To develop and validate a comprehensive framework for obstructive sleep apnea (OSA) detection and forecasting using electrocardiogram (ECG) data.
• To explore the efficacy of integrating dynamical systems theory and statistical analysis for feature engineering in ECG-based OSA detection.
• To implement transfer learning for improved model generalization across datasets and personalization for individual patient data.

Main Methods:

• Utilized two distinct datasets: PhysioNet Apnea-ECG and OSASUD.
• Engineered features from ECG signals using dynamical systems theory and statistical analysis.
• Applied a range of machine learning and deep learning models, incorporating cross-dataset and patient-level transfer learning.

Main Results:

• The developed framework demonstrated effective OSA detection and forecasting capabilities using ECG data.
• Transfer learning strategies successfully improved model generalization to different datasets and enabled personalized predictions.
• The study highlights the potential for ECG-based analysis in precision medicine for OSA management.

Conclusions:

• A novel framework integrating advanced feature engineering and transfer learning for OSA detection from ECG signals has been presented.
• The findings support the use of ECG as a practical tool for OSA screening and diagnosis, paving the way for precision medicine approaches.
• This research offers a scalable and personalized solution for identifying individuals at risk of obstructive sleep apnea.