Event-Level Identification of Sleep Apnea Using FMCW Radar

Figure 1. Overview of the FMCW-radar-based sleep apnea monitoring system.

Figure 2. Schematic diagram of the FMCW radar structure and basic signal-processing flow.

Figure 3. A typical radar respiratory signal from different range bins with a length of 5 min, indicating that the person is experiencing repeated apnea events. From top to bottom are the original IF phase signal, the processed IF phase signal, the synchronized chest and abdominal respiratory banding acquired from PSG, the nasal airflow signal, and the blood oxygen saturation signal.

Figure 4. Range segmentation for bed occupancy detection.

Figure 5. State machine for bed occupancy detection with range-based power-thresholding logic.

Figure 6. The proposed model architecture for respiratory event segmentation and blood oxygen saturation prediction. (a) The overall structure, including the respiratory signal encoder, multihead self-attention bottleneck, respiratory event decoder, and SpO2 prediction decoder, which processes multichannel respiratory signals to output respiratory event segmentation and blood oxygen saturation (SpO2) predictions. (b) The respiratory signal encoder, constructed from four convolutional blocks with downsampling and skip connections. (c) The decoder, also built from four convolutional blocks with upsampling and skip connections. The breath event decoder and SpO2 decoder are based on the same architecture but differ in their output layers.

Figure 7. Histograms and KDE curves for different methods: (a) error in the duration of respiratory events; (b) error in the start time of respiratory events; (c) error in the end time of respiratory events.

Figure 8. Typical signal segments from healthy people (a) and mild apnea (b) and severe apnea (c) patients, from top to bottom, show multichannel-radar-captured respiratory signals, doctor-labeled events based on PSG data, and probabilities/events predicted by the basic U-Net, CNN–MHA, and proposed models.

Figure 9. Confusion matrices, scatter plots, and Bland–Altman analysis graphs for recording-level analyses: (a,d,g) for basic U-Net; (b,e,h) for CNN–MHA; (c,f,i) for this study.

Figure 10. Visualization of multihead self-attention weights revealing normal breathing reference patterns.

Figure 11. Error distribution and cumulative error distribution for SpO2 prediction.

Figure 12. SpO2 predictions with and without the use of correlated losses.