Deep Learning for identifying systolic complexes in SCG traces: a cross-dataset analysis.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Seismocardiography (SCG) offers a non-invasive method for cardiac activity analysis, with the systolic complex being highly informative.
• Current deep learning models for SCG analysis are limited to controlled environments and single datasets, hindering real-world applicability.

Purpose of the Study:

• To evaluate deep learning models for systolic complex detection in a cross-dataset and real-world scenario.
• To investigate the impact of domain shift and personalization on model performance.
• To assess the benefits of a multi-channel approach using accelerometers and gyroscopes.

Main Methods:

• A cross-dataset experimental analysis was conducted using deep learning models.
• Personalization techniques were applied to address domain shift between datasets.
• Multi-channel data from accelerometers and gyroscopes were leveraged.

Main Results:

• Deep learning models demonstrate effectiveness in detecting the seismocardiographic systolic complex.
• A significant domain shift was observed in real-world and cross-dataset scenarios.
• Personalization significantly improved model performance by mitigating the domain shift.
• Multi-channel data fusion enhanced the robustness and accuracy of the analysis.

Conclusions:

• Deep learning is a viable approach for seismocardiogram analysis, but requires adaptation for real-world conditions.
• Personalization is essential for deploying SCG analysis models across different datasets and scenarios.
• Integrating data from multiple sensors (accelerometers and gyroscopes) provides a more comprehensive cardiac signal analysis.