From Motion Artifacts to Clinical Insight: Multi-Modal Deep Learning for Robust Arrhythmia Screening in Ambulatory ECG Monitoring

Figure 1. Complete implementation specification organized by category: infrastructure, data preprocessing, training configuration, regularization strategies, monitoring metrics, and convergence behavior.

Figure 2. Training and validation accuracy over 100 epochs. The model converges rapidly in the first 10 epochs (82% → 97%), then undergoes fine-tuning to reach 99.5% validation accuracy by epoch 30. Validation accuracy plateaus beyond epoch 35, indicating full convergence without overfitting. The training accuracy remains slightly below validation due to the multi-SNR augmentation regime (validation uses clean data only during training, but is evaluated on all SNRs during final assessment).

Figure 3. Arrhythmia detection accuracy versus SNR comparing ECG-only and multi-modal fusion. Accelerometer integration maintains clinically acceptable performance (>85%, dashed line) across all tested conditions, while ECG-only degrades substantially under motion artifacts.

Figure 4. ROC curves stratified by SNR level. Each curve represents classifier performance across all decision thresholds. ECG-only classification (a) shows progressive degradation toward random performance (diagonal line) as noise increases. Multi-modal fusion (b) maintains curves near the ideal upper-left corner, achieving AUC > 0.92 even at −6 dB SNR, where noise power exceeds signal power by 4×.

Figure 5. Direct modality comparison at representative SNR levels (24, 12, 0, −6 dB). (a): ROC space showing ECG + ACC (solid lines) achieves superior true positive rates compared to ECG-only (dashed lines) at matched false positive rates. (b): Precision–recall space demonstrating that fusion maintains precision under noise, whereas ECG-only exhibits rapid precision collapse as recall increases. The gap between methods widens at lower SNRs, confirming accelerometer integration provides the greatest benefit under motion artifact conditions.

Figure 6. Real-world motion artifact effects from the ScientISST MOVE database. (Left): ECG waveform quality progression from sedentary (top) through moderate activity (middle) to vigorous exercise (bottom), showing increasing baseline wander, noise corruption, and morphological distortion. (Right): Corresponding tri-axial accelerometer magnitude demonstrating strong correlation between motion intensity and ECG degradation, validating the motion-artifact coupling assumption underlying sensor fusion.

Figure 7. Correlation between accelerometer magnitude and false positive rates across ScientISST MOVE activities. Scatter plots demonstrate strong linear relationships for both ECG-only (r=0.992, blue squares and line) and ECG + ACC (r=0.982, red circles and line) classifiers, with fusion maintaining consistently lower error rates across the entire motion range from sedentary (0.02 g) to vigorous (0.85 g) activities.

Figure 8. (left) Evolution of precision vs. epochs for the training and validation data; (right) Evolution of precision, recall, and F1 score vs. epochs.