Sleep Apnea Prediction Using Deep Learning.

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Obstructive sleep apnea (OSA) is a prevalent sleep disorder characterized by breathing cessation.
• Current detection methods focus on identifying events as they occur, limiting proactive intervention.
• Predicting OSA events in advance is crucial for developing responsive breathing regulation devices.

Purpose of the Study:

• To develop and evaluate advanced neural network models for predicting obstructive sleep apnea (OSA) events.
• To assess the efficacy of using raw respiratory signals without engineered features for OSA prediction.
• To determine the feasibility of real-time, advance prediction of OSA events for potential therapeutic devices.

Main Methods:

• Four deep learning models were proposed: 1D-CNN, ConvLSTM, 1D-CNN-LSTM, and 2D-CNN-LSTM.
• Models utilized raw nasal flow, abdominal, and thoracic respiratory signals sampled at 32 Hz.
• Prediction of OSA events (apnea/hypopnea) and normal breathing was performed 30 seconds ahead using 90 seconds of prior data.

Main Results:

• All four models demonstrated promising accuracy exceeding 81% on a large dataset (1,507 subjects, >46,000 examples).
• 1D-CNN-LSTM and 2D-CNN-LSTM models achieved the highest performance with accuracy, sensitivity, and specificity over 83%, 81%, and 85%, respectively.
• The 1D-CNN-LSTM model maintained high performance (82.94% accuracy) even with signals downsampled to 1 Hz, indicating robustness.

Conclusions:

• Accurate, advance prediction of OSA events is achievable using deep learning models and raw respiratory signals.
• The proposed models, particularly 1D-CNN-LSTM and 2D-CNN-LSTM, offer a viable approach for developing preemptive OSA management devices.
• The robustness to low sampling frequencies makes these algorithms suitable for low-resource, at-home monitoring devices.