MPCNN: A Novel Matrix Profile Approach for CNN-based Single Lead Sleep Apnea in Classification Problem.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Sleep apnea (SA) is a critical global health issue.
• Deep Learning (DL) shows promise for electrocardiogram (ECG)-based SA diagnosis.
• Conventional ECG features (R-peaks, RR intervals) may miss vital signal information.

Purpose of the Study:

• To develop an innovative ECG feature extraction method for improved SA detection.
• To address the limitations of traditional feature extraction in DL models for SA diagnosis.
• To enhance the performance of lightweight DL models for SA detection.

Main Methods:

• Inspired by Matrix Profile algorithms, derived MinDP, MaxDP, and MeanDP from ECG signal subsequences.
• Utilized modified LeNet-5, BAFNet, and SE-MSCNN deep learning models.
• Validated the approach on PhysioNet Apnea-ECG and UCDDB datasets.

Main Results:

• Achieved up to 92.11% and 81.25% per-segment accuracy on the datasets.
• Reached 100% per-recording accuracy on the PhysioNet data.
• Obtained a high correlation of 0.989 with state-of-the-art methods.

Conclusions:

• The novel distance-based feature extraction method enhances DL model performance for SA detection.
• The approach shows potential for application in home sleep apnea tests (HSAT) and IoT devices.
• Publicly available source code facilitates further research and development.