Multiscale Bidirectional Temporal Convolutional Network for Sleep Apnea Detection Based on Wearable...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Sleep Medicine

Background:

• Sleep apnea syndrome (SAS) is a prevalent chronic disorder linked to cardiopulmonary diseases.
• Early and accessible diagnosis of SAS is crucial for timely intervention.
• Wearable devices offer a promising avenue for unobtrusive, continuous health monitoring.

Purpose of the Study:

• To develop and validate an advanced AI-driven method for the unobtrusive detection of sleep apnea syndrome (SAS).
• To leverage photoplethysmography (PPG) data from wearable devices for SAS identification.
• To enhance model robustness and address data challenges like overfitting and class imbalance.

Main Methods:

• Implementation of a one-dimensional multi-scale bidirectional temporal convolutional neural network (1D-MsBiTCNet) for feature extraction and temporal analysis.
• Application of regularized dropout (RD) to mitigate model overfitting during training.
• Utilization of logit adjustment (LA) to address class imbalance in the dataset.

Main Results:

• The 1D-MsBiTCNet model achieved high performance metrics on a PPG dataset from 92 subjects.
• Per-segment detection yielded accuracy (82.76%), sensitivity (71.58%), and specificity (86.74%).
• Per-recording severe SAS detection reached accuracy (97.83%), sensitivity (88.89%), and specificity (100.00%), with a mean absolute error of 5.44 for AHI prediction.

Conclusions:

• The proposed 1D-MsBiTCNet method demonstrates outstanding performance for SAS detection using wearable PPG data.
• RD and LA techniques effectively improved model training and addressed dataset limitations.
• This approach offers a valuable methodological reference for large-scale, automated SAS detection systems.