SCG systolic detection in the wild: A static-dynamic cross-dataset analysis.

Area of Science:

• Biomedical Engineering
• Cardiovascular Monitoring
• Signal Processing

Background:

• Seismocardiography (SCG) offers continuous monitoring of cardiac electromechanical activity.
• Accurate detection of the systolic complex within SCG signals is essential for clinical analysis.
• Current Deep Learning (DL) methods for SCG analysis are often evaluated in controlled settings, limiting real-world applicability.

Purpose of the Study:

• To conduct an exhaustive experimental analysis of SCG systolic complex detection in real-world, dynamic conditions.
• To quantify the impact of domain shift on DL model performance for SCG analysis.
• To evaluate the effectiveness of model adaptation strategies and multi-channel approaches in challenging environments.

Main Methods:

• Utilized a U-Net based Deep Learning model for systolic complex detection.
• Designed a novel experimental framework combining multiple SCG datasets.
• Performed cross-dataset and static-dynamic domain shift evaluations to simulate real-world variability.
• Assessed traditional model adaptation techniques like fine-tuning and personalization.
• Investigated the benefits of a multi-channel approach.

Main Results:

• Real-world SCG detection is significantly more challenging than in controlled settings.
• Both cross-dataset and static-dynamic domain shifts notably degrade detection performance.
• Traditional adaptation strategies offered only partial improvements in performance.
• A multi-channel approach demonstrated significant benefits in real-world SCG detection scenarios.

Conclusions:

• Deep learning models for SCG systolic detection face substantial challenges in dynamic, real-world environments due to domain shifts.
• Model adaptation strategies require further development to effectively address real-world variability.
• Multi-channel SCG acquisition shows promise for improving the robustness of cardiac monitoring systems.