Deep Learning-Based Heartbeat Detection from 3D Seismocardiography for Robust Heart Rate Monitoring

Figures

• 

  Figure A1. Examples of model outputs under diverse SCG scenarios: (a) high-quality signal, (b) noisy signal with successful detection, (c) motion-affected segment, (d) signals where beats are not visible, and (e) signal where heartbeats exhibit a double-peaked waveform, causing duplicate detections.

• 

  Figure 1. Partitioning of the resting SCG database into training, validation, and test sets with subject and segment counts.

• 

  Figure 2. Preprocessing and heartbeat annotation of SCG signals within a 5-s window. The SCG signal is shown in blue, and the rectangular heartbeat mask is shown in red.

• 

  Figure 3. Architecture of the 1D U-Net model designed for tri-axial SCG input. The network processes 3D SCG signals and outputs a 1D mask indicating R-peak locations.

• 

  Figure 4. Heartbeat detection performance metrics (PPV, Sensitivity, and F1-score) across all subjects before (solid line) and after (dashed line) movement artifact correction.

• 

  Figure 5. Bland–Altman analysis for HR estimation illustrating the agreement between estimated and reference HR with mean bias and standard error of estimation (SEE); outliers (in red) are clipped at ±10 bpm for visualization. The middle dashed line represents the mean bias, while the upper and lower dashed lines represent the limits of agreement (mean ± SD).

• 

  Figure 6. Z-axis acceleration signal from a representative nocturnal recording together with the corresponding reference HR and the estimated HR, demonstrating their agreement over time.